Enterprise data teams moving agentic AI into production are hitting a consistent failure point at the data tier. Agents built across a vector store, a relational database, a graph store and a lakehouse require sync pipelines to keep context current. Under production load, that context goes stale.
Oracle, whose database infrastructure runs the transaction systems of 97% of Fortune Global 100 companies by the company’s own count, is now making a direct architectural argument that the database is the right place to fix that problem.
Oracle this week announced a set of agentic AI capabilities for Oracle AI Database, built around a direct architectural counter-argument to that pattern.
The core of the release is the Unified Memory Core, a single ACID (Atomicity, Consistency, Isolation, and Durability)-transactional engine that processes vector, JSON, graph, relational, spatial and columnar data without a sync layer. Alongside that, Oracle announced Vectors on Ice for native vector indexing on Apache Iceberg tables, a standalone Autonomous AI Vector Database service and an Autonomous AI Database MCP Server for direct agent access without custom integration code.
The news isn’t just that Oracle is adding new features, it’s about the world’s largest database vendor realizing that things have changed in the AI world that go beyond what its namesake database was providing.
“As much as I’d love to tell you that everybody stores all their data in an Oracle database today — you and I live in the real world,” Maria Colgan, Vice President, Product Management for Mission-Critical Data and AI Engines, at Oracle told VentureBeat. “We know that that’s not true.”
Oracle’s release spans four interconnected capabilities. Together they form the architectural argument that a converged database engine is a better foundation for production agentic AI than a stack of specialized tools.
Unified Memory Core. Agents reasoning across multiple data formats simultaneously — vector, JSON, graph, relational, spatial — require sync pipelines when those formats live in separate systems. The Unified Memory Core puts all of them in a single ACID-transactional engine. Under the hood it is an API layer over the Oracle database engine, meaning ACID consistency applies across every data type without a separate consistency mechanism.
“By having the memory live in the same place that the data does, we can control what it has access to the same way we would control the data inside the database,” Colgan explained.
Vectors on Ice. For teams running data lakehouse architectures on the open-source Apache Iceberg table format, Oracle now creates a vector index inside the database that references the Iceberg table directly. The index updates automatically as the underlying data changes and works with Iceberg tables that are managed by Databricks and Snowflake. Teams can combine Iceberg vector search with relational, JSON, spatial or graph data stored inside Oracle in a single query.
Autonomous AI Vector Database. A fully managed, free-to-start vector database service built on the Oracle 26ai engine. The service is designed as a developer entry point with a one-click upgrade path to full Autonomous AI Database when workload requirements grow.
Autonomous AI Database MCP Server. Lets external agents and MCP clients connect to Autonomous AI Database without custom integration code. Oracle’s row-level and column-level access controls apply automatically when an agent connects, regardless of what the agent requests.
“Even though you are making the same standard API call you would make with other platforms, the privileges that user has continued to kick in when the LLM is asking those questions,” Colgan said.
Oracle’s Autonomous AI Vector Database enters a market occupied by purpose-built vector services including Pinecone, Qdrant and Weaviate. The distinction Oracle is drawing is about what happens when vector alone is not enough.
“Once you are done with vectors, you do not really have an option,” Steve Zivanic, Global Vice President, Database and Autonomous Services, Product Marketing at Oracle, told VentureBeat. “With this, you can get graph, spatial, time series — whatever you may need. It is not a dead end.”
Holger Mueller, principal analyst at Constellation Research, said that the architectural argument is credible precisely because other vendors cannot make it without moving data first. Other database vendors require transactional data to move to a data lake before agents can reason across it. Oracle’s converged legacy, in his view, gives it a structural advantage that is difficult to replicate without a ground-up rebuild.
Not everyone sees the feature set as differentiated. Steven Dickens, CEO and principal analyst at HyperFRAME Research, told VentureBeat that vector search, RAG integration and Apache Iceberg support are now standard requirements across enterprise databases — Postgres, Snowflake and Databricks all offer comparable capabilities.
“Oracle’s move to label the database itself as an AI Database is primarily a rebranding of its converged database strategy to match the current hype cycle,” Dickens said. In his view the real differentiation Oracle is claiming is not at the feature level but at the architectural level — and the Unified Memory Core is where that argument either holds or falls apart.
The four capabilities Oracle shipped this week are a response to a specific and well-documented production failure mode. Enterprise agent deployments are not breaking down at the model layer. They are breaking down at the data layer, where agents built across fragmented systems hit sync latency, stale context and inconsistent access controls the moment workloads scale.
Matt Kimball, vice president and principal analyst at Moor Insights and Strategy, told VentureBeat the data layer is where production constraints surface first.
“The struggle is running them in production,” Kimball said. “The gap is seen almost immediately at the data layer — access, governance, latency and consistency. These all become constraints.”
Dickens frames the core mismatch as a stateless-versus-stateful problem. Most enterprise agent frameworks store memory as a flat list of past interactions, which means agents are effectively stateless while the databases they query are stateful. The lag between the two is where decisions go wrong.
“Data teams are exhausted by fragmentation fatigue,” Dickens said. “Managing a separate vector store, graph database and relational system just to power one agent is a DevOps nightmare.”
That fragmentation is precisely what Oracle’s Unified Memory Core is designed to eliminate. The control plane question follows directly.
“In a traditional application model, control lives in the app layer,” Kimball said. “With agentic systems, access control breaks down pretty quickly because agents generate actions dynamically and need consistent enforcement of policy. By pushing all that control into the database, it can all be applied in a more uniform way.”
The question of where control lives in an enterprise agentic AI stack is not settled.
Most organizations are still building across fragmented systems, and the architectural decisions being made now — which engine anchors agent memory, where access controls are enforced, how lakehouse data gets pulled into agent context — will be difficult to undo at scale.
The distributed data challenge is still the real test.
“Data is increasingly distributed across SaaS platforms, lakehouses and event-driven systems, each with its own control plane and governance model,” Kimball said. “The opportunity now is extending that model across the broader, more distributed data estates that define most enterprise environments today.”
Engineers building browser agents today face a choice between closed APIs they cannot inspect and open-weight frameworks with no trained model underneath them. Ai2 is now offering a third option.
The Seattle-based nonprofit behind the open-source OLMo language models and Molmo vision-language family today is releasing MolmoWeb, an open-weight visual web agent available in 4 billion and 8 billion parameter sizes.
Until now, no open-weight visual web agent shipped with the training data and pipeline needed to audit or reproduce it. MolmoWeb does.
MolmoWebMix, the accompanying dataset, includes 30,000 human task trajectories across more than 1,100 websites, 590,000 individual subtask demonstrations and 2.2 million screenshot question-answer pairs — which Ai2 describes as the largest publicly released collection of human web-task execution ever assembled.
“Can you go from just passively understanding images, describing them and captioning them, to actually making them take action in some environment?” Tanmay Gupta, senior research scientist at Ai2, told VentureBeat. “That is exactly what MolmoWeb is.”
MolmoWeb operates entirely from browser screenshots. It does not parse HTML or rely on accessibility tree representations of a page. At each step it receives a task instruction, the current screenshot, a text log of previous actions and the current URL and page title. It produces a natural-language thought describing its reasoning, then executes the next browser action — clicking at screen coordinates, typing text, scrolling, navigating to a URL or switching tabs.
The model is browser-agnostic. It requires only a screenshot, which means it runs against local Chrome, Safari or a hosted browser service. The hosted demo uses Browserbase, a cloud browser infrastructure startup.
The model weights are only part of what Ai2 is releasing. MolmoWebMix, the accompanying training dataset, is the core differentiator from every other open-weight agent available today.
“The data basically looks like a sequence of screenshots and actions paired with instructions for what the intent behind that sequence of screenshots was,” Gupta said.
MolmoWebMix combines three components.
Human demonstrations. Human annotators completed browsing tasks using a custom Chrome extension that recorded actions and screenshots across more than 1,100 websites. The result is 30,000 task trajectories spanning more than 590,000 individual subtask demonstrations.
Synthetic trajectories. To scale beyond what human annotation alone can provide, Ai2 generated additional trajectories using text-based accessibility-tree agents — single-agent runs filtered for task success, multi-agent pipelines that decompose tasks into subgoals and deterministic navigation paths across hundreds of websites. Critically, no proprietary vision agents were used. The synthetic data came from text-only systems, not from OpenAI Operator or Anthropic’s computer use API.
GUI perception data. A third component trains the model to read and reason about page content directly from images. It includes more than 2.2 million screenshot question-answer pairs drawn from nearly 400 websites, covering element grounding and screenshot-based reasoning tasks.
“If you are able to perform a task and you’re able to record a trajectory from that, you should be able to train the web agent on that trajectory to do the exact same task,” Gupta said.
In Gupta’s view, there are two categories of technologies in the browser agent market.
The first is API-only systems, capable but closed, with no visibility into training or architecture. OpenAI Operator, Anthropic’s computer use API and Google’s Gemini computer use fall into this group.
The second is open-weight models, a significantly smaller category. Browser-use, the most widely adopted open alternative, is a framework rather than a trained model. It requires developers to supply their own LLM and build the agent layer on top.
MolmoWeb sits in the second category as a fully trained open-weight vision model. Ai2 reports it leads that group across four live-website benchmarks: WebVoyager, Online-Mind2Web, DeepShop and WebTailBench. According to Ai2, it also outperforms older API-based agents built on GPT-4o with accessibility tree plus screenshot input.
Ai2 documents several current limitations in the release. The model makes occasional errors reading text from screenshots, drag-and-drop interactions remain unreliable and performance degrades on ambiguous or heavily constrained instructions. The model was also not trained on tasks requiring logins or financial transactions.
Enterprise teams evaluating browser agents are not just choosing a model. They are deciding whether they can audit what they are running, fine-tune it on internal workflows, and avoid a per-call API dependency.
Getting AI agents to perform reliably in production — not just in demos — is turning out to be harder than enterprises anticipated. Fragmented data, unclear workflows, and runaway escalation rates are slowing deployments across industries.
“The technology itself often works well in demonstrations,” said Sanchit Vir Gogia, chief analyst with Greyhound Research. “The challenge begins when it is asked to operate inside the complexity of a real organization.”
Burley Kawasaki, who oversees agent deployment at Creatio, and team have developed a methodology built around three disciplines: data virtualization to work around data lake delays; agent dashboards and KPIs as a management layer; and tightly bounded use-case loops to drive toward high autonomy.
In simpler use cases, Kawasaki says these practices have enabled agents to handle up to 80-90% of tasks on their own. With further tuning, he estimates they could support autonomous resolution in at least half of use cases, even in more complex deployments.
“People have been experimenting a lot with proof of concepts, they’ve been putting a lot of tests out there,” Kawasaki told VentureBeat. “But now in 2026, we’re starting to focus on mission-critical workflows that drive either operational efficiencies or additional revenue.”
Enterprises are eager to adopt agentic AI in some form or another — often because they’re afraid to be left out, even before they even identify real-world tangible use cases — but run into significant bottlenecks around data architecture, integration, monitoring, security, and workflow design.
The first obstacle almost always has to do with data, Gogia said. Enterprise information rarely exists in a neat or unified form; it is spread across SaaS platforms, apps, internal databases, and other data stores. Some are structured, some are not.
But even when enterprises overcome the data retrieval problem, integration is a big challenge. Agents rely on APIs and automation hooks to interact with applications, but many enterprise systems were designed long before this kind of autonomous interaction was a reality, Gogia pointed out.
This can result in incomplete or inconsistent APIs, and systems can respond unpredictably when accessed programmatically. Organizations also run into snags when they attempt to automate processes that were never formally defined, Gogia said.
“Many business workflows depend on tacit knowledge,” he said. That is, employees know how to resolve exceptions they’ve seen before without explicit instructions — but, those missing rules and instructions become startlingly obvious when workflows are translated into automation logic.
Creatio deploys agents in a “bounded scope with clear guardrails,” followed by an “explicit” tuning and validation phase, Kawasaki explained. Teams review initial outcomes, adjust as needed, then re-test until they’ve reached an acceptable level of accuracy.
That loop typically follows this pattern:
Design-time tuning (before go-live): Performance is improved through prompt engineering, context wrapping, role definitions, workflow design, and grounding in data and documents.
Human-in-the-loop correction (during execution): Devs approve, edit, or resolve exceptions. In instances where humans have to intervene the most (escalation or approval), users establish stronger rules, provide more context, and update workflow steps; or, they’ll narrow tool access.
Ongoing optimization (after go-live): Devs continue to monitor exception rates and outcomes, then tune repeatedly as needed, helping to improve accuracy and autonomy over time.
Kawasaki’s team applies retrieval-augmented generation to ground agents in enterprise knowledge bases, CRM data, and other proprietary sources.
Once agents are deployed in the wild, they are monitored with a dashboard providing performance analytics, conversion insights, and auditability. Essentially, agents are treated like digital workers. They have their own management layer with dashboards and KPIs.
For instance, an onboarding agent will be incorporated as a standard dashboard interface providing agent monitoring and telemetry. This is part of the platform layer — orchestration, governance, security, workflow execution, monitoring, and UI embedding — that sits “above the LLM,” Kawasaki said.
Users see a dashboard of agents in use and each of their processes, workflows, and executed results. They can “drill down” into an individual record (like a referral or renewal) that shows a step-by-step execution log and related communications to support traceability, debugging, and agent tweaking. The most common adjustments involve logic and incentives, business rules, prompt context, and tool access, Kawasaki said.
The biggest issues that come up post-deployment:
Exception handling volume can be high: Early spikes in edge cases often occur until guardrails and workflows are tuned.
Data quality and completeness: Missing or inconsistent fields and documents can cause escalations; teams can identify which data to prioritize for grounding and which checks to automate.
Auditability and trust: Regulated customers, particularly, require clear logs, approvals, role-based access control (RBAC), and audit trails.
“We always explain that you have to allocate time to train agents,” Creatio’s CEO Katherine Kostereva told VentureBeat. “It doesn’t happen immediately when you switch on the agent, it needs time to understand fully, then the number of mistakes will decrease.”
When looking to deploy agents, “Is my data ready?,” is a common early question. Enterprises know data access is important, but can be turned off by a massive data consolidation project.
But virtual connections can allow agents access to underlying systems and get around typical data lake/lakehouse/warehouse delays. Kawasaki’s team built a platform that integrates with data, and is now working on an approach that will pull data into a virtual object, process it, and use it like a standard object for UIs and workflows. This way, they don’t have to “persist or duplicate” large volumes of data in their database.
This technique can be helpful in areas like banking, where transaction volumes are simply too large to copy into CRM, but are “still valuable for AI analysis and triggers,” Kawasaki said.
Once integrations and virtual objects are established, teams can evaluate data completeness, consistency, and availability, and identify low-friction starting points (like document-heavy or unstructured workflows).
Kawasaki emphasized the importance of “really using the data in the underlying systems, which tends to actually be the cleanest or the source of truth anyway.”
The best fit for autonomous (or near-autonomous) agents are high-volume workflows with “clear structure and controllable risk,” Kawasaki said. For instance, document intake and validation in onboarding or loan preparation, or standardized outreach like renewals and referrals.
“Especially when you can link them to very specific processes inside an industry — that’s where you can really measure and deliver hard ROI,” he said.
For instance, financial institutions are often siloed by nature. Commercial lending teams perform in their own environment, wealth management in another. But an autonomous agent can look across departments and separate data stores to identify, for instance, commercial customers who might be good candidates for wealth management or advisory services.
“You think it would be an obvious opportunity, but no one is looking across all the silos,” Kawasaki said. Some banks that have applied agents to this very scenario have seen “benefits of millions of dollars of incremental revenue,” he claimed, without naming specific institutions.
However, in other cases — particularly in regulated industries — longer-context agents are not only preferable, but necessary. For instance, in multi-step tasks like gathering evidence across systems, summarizing, comparing, drafting communications, and producing auditable rationales.
“The agent isn’t giving you a response immediately,” Kawasaki said. “It may take hours, days, to complete full end-to-end tasks.”
This requires orchestrated agentic execution rather than a “single giant prompt,” he said. This approach breaks work down into deterministic steps to be performed by sub-agents. Memory and context management can be maintained across various steps and time intervals. Grounding with RAG can help keep outputs tied to approved sources, and users have the ability to dictate expansion to file shares and other document repositories.
This model typically doesn’t require custom retraining or a new foundation model. Whatever model enterprises use (GPT, Claude, Gemini), performance improves through prompts, role definitions, controlled tools, workflows, and data grounding, Kawasaki said.
The feedback loop puts “extra emphasis” on intermediate checkpoints, he said. Humans review intermediate artifacts (such as summaries, extracted facts, or draft recommendations) and correct errors. Those can then be converted into better rules and retrieval sources, narrower tool scopes, and improved templates.
“What is important for this style of autonomous agent, is you mix the best of both worlds: The dynamic reasoning of AI, with the control and power of true orchestration,” Kawasaki said.
Ultimately, agents require coordinated changes across enterprise architecture, new orchestration frameworks, and explicit access controls, Gogia said. Agents must be assigned identities to restrict their privileges and keep them within bounds. Observability is critical; monitoring tools can record task completion rates, escalation events, system interactions, and error patterns. This kind of evaluation must be a permanent practice, and agents should be tested to see how they react when encountering new scenarios and unusual inputs.
“The moment an AI system can take action, enterprises have to answer several questions that rarely appear during copilot deployments,” Gogia said. Such as: What systems is the agent allowed to access? What types of actions can it perform without approval? Which activities must always require a human decision? How will every action be recorded and reviewed?
“Those [enterprises] that underestimate the challenge often find themselves stuck in demonstrations that look impressive but cannot survive real operational complexity,” Gogia said.
Voice AI is moving faster than the tools we use to measure it. Every major AI lab — OpenAI, Google DeepMind, Anthropic, xAI — is racing to ship voice models capable of natural, real-time conversation.
But the benchmarks used to evaluate those models are largely still running on synthetic speech, English-only prompts, and scripted test sets that bear little resemblance to how people actually talk.
Scale AI, the large data annotation startup whose founder was poached by Meta last year to lead its Superintelligence Lab, is still going strong and tackling the problem head on: today it launches Voice Showdown, what it calls the first global preference-based arena designed to benchmark voice AI through the lens of real human interaction.
This product offers a unique strategic value to users: free access to the world’s leading frontier models. Through Scale’s ChatLab platform, users can interact with high-tier models—which typically require multiple $20-per-month subscriptions—at no cost. In exchange, users participate in occasional blind, head-to-head “battles” to choose which of two anonymized leading voice models offers a better experience, providing data for the industry’s most authentic, human-preference leaderboard of voice AI models.
“Voice AI is really the fastest moving frontier in AI right now,” said Janie Gu, product manager for Showdown at Scale AI. “But the way that we evaluate voice models hasn’t kept up.”
The results, drawn from thousands of spontaneous voice conversations across more than 60 languages, reveal capability gaps that other benchmarks have consistently missed.
Voice Showdown is built on ChatLab, Scale’s model-agnostic chat platform where users can freely interact with whichever frontier AI model they choose — for free — within a single app. The platform has been available to Scale’s global community of over 500,000 annotators, with roughly 300,000 having submitted at least one prompt. Scale is opening the platform to a public waitlist today.
The evaluation mechanism is elegant in its simplicity: while a user is having a natural voice conversation with a model, the system occasionally — on fewer than 5% of all voice prompts — surfaces a blind side-by-side comparison. The same prompt is sent to a second, anonymous model, and the user picks which response they prefer.
This design solves three problems that plague existing voice benchmarks.
First, every prompt comes from real human speech — with accents, background noise, half-finished sentences, and conversational filler — rather than synthesized audio generated from text.
Second, the platform spans more than 60 languages across 6 continents, with over a third of battles occurring in non-English languages including Spanish, Arabic, Japanese, Portuguese, Hindi, and French.
Third, because battles occur within users’ actual daily conversations, 81% of prompts are conversational or open-ended — questions without a single correct answer. That rules out automated scoring and makes human preference the only credible signal.
Voice Showdown currently runs two evaluation modes: Dictate (users speak, models respond with text) and Speech-to-Speech, or S2S (Speech-to-Speech, users speak, models talk back). A third mode — Full Duplex, which captures real-time, interruptible conversation — is in development.
One design detail sets Voice Showdown apart from Chatbot Arena (LM Arena), the text benchmark it most closely resembles. In LM Arena, critics have noted that users sometimes cast throwaway votes with little stake in the outcome. Voice Showdown addresses this directly: after a user votes for the model they preferred, the app switches them to that model for the rest of their conversation. If you voted for GPT-4o Audio over Gemini, you’re now talking to GPT-4o Audio. That alignment of consequence with preference discourages casual or dishonest voting.
The system also controls for confounds that could corrupt comparisons: both model responses begin streaming simultaneously (eliminating speed bias), voice gender is matched across both options (eliminating gender preference bias), and neither model is identified by name during voting.
Voice Showdown launches with 11 frontier models evaluated across 52 model-voice pairs as of March 18, 2026. Not all models support both evaluation modes — the Dictate leaderboard includes 8 models, while S2S includes 6.
Dictate Leaderboard (Speech-In, Text-Out)
In this mode, users provide a spoken prompt and evaluate two side-by-side text responses. Here are the baseline scores:
Gemini 3 Pro (1073)
Gemini 3 Flash (1068)
GPT-4o Audio (1019)
Qwen 3 Omni (1000)
Voxtral Small (925)
Gemma 3n (918)
GPT Realtime (875)
Phi-4 Multimodal (729)
Note: Gemini 3 Pro and Gemini 3 Flash are statistically tied for the top rank.
Speech-to-Speech (S2S) Leaderboard
In this mode, users speak to the model and evaluate two competing audio responses. Also baselines:
Gemini 2.5 Flash Audio (1060)
GPT-4o Audio (1059)
Grok Voice (1024)
Qwen 3 Omni (1000)
GPT Realtime (962)
GPT Realtime 1.5 (920)
Note: Gemini 2.5 Flash Audio and GPT-4o Audio are statistically tied for the top rank in baseline evaluations.
Dictate rankings are led by Google’s Gemini 3 Pro and Gemini 3 Flash, which are statistically tied at #1 with Elo scores around 1,043-1,044 after style controls.
GPT-4o Audio holds a clear third place. Open-weight models including Gemma3n, Voxtral Small, and Phi-4 Multimodal trail significantly.
Speech-to-Speech (S2S) rankings show a tighter race at the top, with Gemini 2.5 Flash Audio and GPT-4o Audio statistically tied at #1 in the baseline rankings.
After adjusting for response length and formatting — factors that can inflate perceived quality — GPT-4o Audio pulls ahead (1,102 Elo vs. 1,075 for Gemini 2.5 Flash Audio).
Grok Voice jumps to a close second at 1,093 under style controls, suggesting its raw #3 ranking undersells its actual performance quality.
Qwen 3 Omni, the open-weight model from Alibaba’s Qwen team, performs better on pure preference than its popularity would suggest — ranking fourth in both modes, ahead of several higher-profile names.
“When people come in, they go for the big names,” Gu noted. “But for preference, lesser-known models like Qwen actually pull ahead.”
Beyond rankings, Voice Showdown’s real value is in the failure diagnostics — and those paint a more complicated picture of voice AI than most leaderboards reveal.
The multilingual gap is worse than you think
Language robustness is the starkest differentiator across models. In Dictate, Gemini 3 models lead across essentially every language tested.
In S2S, the winner depends heavily on which language is being spoken: GPT-4o Audio leads in Arabic and Turkish; Gemini 2.5 Flash Audio is strongest in French; Grok Voice is competitive in Japanese and Portuguese.
But the more alarming finding is how frequently some models simply stop responding in the user’s language at all.
GPT Realtime 1.5 — OpenAI’s newer real-time voice model — responds in English to non-English prompts roughly 20% of the time, even on high-resource, officially supported languages like Hindi, Spanish, and Turkish.
Its predecessor, GPT Realtime, mismatches at about half that rate (~10%). Gemini 2.5 Flash Audio and GPT-4o Audio sit at ~7%.
The phenomenon runs both directions: some models carry non-English context from earlier in a conversation into an English turn, or simply mishear a prompt and generate an unrelated response in the wrong language entirely.
User verbatims from the platform capture the frustration bluntly: “I said I have an interview today with Quest Management and instead of answering, it gave me information about ‘Risk Management.'”
“GPT Realtime 1.5 thought I was speaking incoherently and recommended mental health assistance, while Qwen 3 Omni correctly identified I was speaking a Nigerian local language.”
The reason existing benchmarks miss this: they’re built on synthetic speech optimized for clean acoustic conditions, and they’re rarely multilingual. Real speakers in real environments — with background noise, short utterances, and regional accents — break speech understanding in ways lab conditions don’t anticipate.
Voice Showdown evaluates models not just at the model level but at the individual voice level — and the variance within a single model’s voice catalog is striking.
For one unnamed model in the study, the best-performing voice won 30 percentage points more often than the worst-performing voice from the same underlying model. Both voices share the same reasoning and generation backend. The difference is purely in audio presentation.
The top-performing voices tend to win or lose on audio understanding and content completeness — whether the model heard you correctly and answered fully. But speech quality remains a deciding factor at the voice selection level, particularly when models are otherwise comparable. “Voice directly shapes how users evaluate the interaction,” Gu said.
Most benchmarks test a single turn. Voice Showdown tests how models hold up across extended conversations — and the results aren’t flattering.
On Turn 1, content quality accounts for 23% of model failures. By Turn 11 and beyond, it becomes the primary failure mode at 43%. Most models see their win rates decline as conversations extend, struggling to maintain coherence across multiple exchanges.
GPT Realtime variants are an exception, marginally improving on later turns — consistent with their known strengths on longer contexts, and their documented weakness on the brief, noisy utterances that dominate early interactions.
Prompt length shows a complementary pattern: short prompts (under 10 seconds) are dominated by audio understanding failures (38%), while long prompts (over 40 seconds) shift the primary failure toward content quality (31%). Shorter audio gives models less acoustic context to parse; longer requests are understood but harder to answer well.
After every S2S comparison, users tag why they preferred one response over the other across three axes: audio understanding, content quality, and speech output. The failure signatures differ meaningfully by model.
Qwen 3 Omni’s losses cluster around speech generation — its reasoning is competitive, but users are put off by how it sounds. GPT Realtime 1.5’s losses are dominated by audio understanding failures (51%), consistent with its language-switching behavior on challenging prompts. Grok Voice’s failures are more balanced across all three axes, indicating no single dominant weakness but no particular strength either.
The current leaderboard covers turn-based interaction — you speak, the model responds, repeat. But real voice conversations don’t work that way. People interrupt, change direction mid-sentence, and talk over each other.
Scale says Full Duplex evaluation — designed to capture these real-time dynamics through human preference rather than scripted scenarios or automated metrics — is coming to Showdown next. No existing benchmark captures full-duplex interaction through organic human preference data.
The leaderboard is live at scale.com/showdown. A public waitlist to join ChatLab and vote on comparisons is open today, with users receiving free access to frontier voice models including GPT-4o, Gemini, and Grok in exchange for occasional preference votes.
In 2026, data engineers working with multi-agent systems are hitting a familiar problem: Agents built on different platforms don’t operate from a shared understanding of the business. The result isn’t model failure — it’s hallucination driven by fragmented context.
The problem is that agents built on different platforms, by different teams, do not share a common understanding of how the business actually operates. Each one carries its own interpretation of what a customer, an order or a region means. When those definitions diverge across a workforce of agents, decisions break down.
A set of announcements from Microsoft this week directly targets that problem. The centerpiece is a significant expansion of Fabric IQ, the semantic intelligence layer the company debuted in November 2025. Fabric IQ’s business ontology is now accessible via MCP to any agent from any vendor, not just Microsoft’s. Alongside that, Microsoft is adding enterprise planning to Fabric IQ, unifying historical data, real-time signals and formal organizational goals in one queryable layer. The new Database Hub brings Azure SQL, Cosmos DB, PostgreSQL, MySQL and SQL Server under a single management plane inside Fabric. Fabric data agents reach general availability.
The overall goal is a unified platform where all data and semantics are available and accessible by any agent to get the context that enterprises require.
Amir Netz, CTO of Microsoft Fabric, reached for a film analogy to explain why the shared context layer matters. “It’s a little bit like the girl from 50 First Dates,” Netz told VentureBeat. “Every morning they wake up and they forget everything and you have to explain it again. This is the explanation that you give them every morning.”
Making the ontology MCP-accessible is the step that moves Fabric IQ from a Fabric-specific feature into shared infrastructure for multi-vendor agent deployments. Netz was explicit about the design intent.
“It doesn’t really matter whose agent it is, how it was built, what the role is,” Netz said. “There’s certain common knowledge, certain common context that all the agents will share.”
That shared context is also where Netz draws a clear line between what the ontology does and what RAG does. He did not dismiss retrieval-augmented generation as a technique — he placed it specifically. RAG handles large document bodies such as regulations, company handbooks and technical documentation, where on-demand retrieval is more practical than loading everything into context.
“We don’t expect humans to remember everything by heart,” he said. “When somebody asks a question, you have to know to go and do a little bit of a search, find the right relevant part and bring it back.”
But RAG does not solve for real-time business state, he argued. It does not tell an agent which planes are in the air right now, whether a crew has enough rest hours, or what the current priority is on a given product line.
“The mistake of the past was they thought one technology can just give you everything,” Netz said. “The cognitive model of the agents is similar to humans. You have to have things that are available out of memory, things that are available on demand, things that are constantly observed and detected in real time.”
Industry analysts see the logic behind Microsoft’s direction but have questions about what comes next.
Robert Kramer, analyst at Moor Insights and Strategy, noted that Microsoft’s broad stack gives it a structural advantage in the race to become the default platform for enterprise agent deployments.
“Fabric ties into Power BI, Microsoft 365, Dynamics and Azure services. That gives Microsoft a natural path to connect enterprise data with business users, operational workflows and now AI systems operating across that environment,” he said. The trade-off, Kramer said, is that Microsoft is competing across a wider surface area than Databricks or Snowflake, which built their reputations on depth of the data platform itself.
The more immediate question for data teams, Kramer said, is whether MCP access actually reduces integration work.
“Most enterprises do not operate in a single AI environment. Finance might be using one set of tools, engineering another, supply chain something else,” Kramer told VentureBeat. “If Fabric IQ can act as a common data context layer those agents can access, it starts to reduce some of the fragmentation that typically shows up around enterprise data.”
But, he said, “If it just adds another protocol that still requires a lot of engineering work, adoption will be slower.”
Whether the engineering work is the harder problem is open to debate. Independent analyst Sanjeev Mohan, told VentureBeat, that the bigger challenge is organizational, not technical.
“I don’t think they fully understand the implications yet,” he said of enterprise data teams. “This is a classical capabilities overhang — capabilities are expanding faster than people’s imagination to use them. The harder work will be ensuring that the context layer is reliable and trustworthy.”
Holger Mueller, principal analyst at Constellation Research, sees MCP as the right mechanism but urges caution on execution.
“For enterprise to benefit from AI, they need to get access to their data — that is in many places unorganized, siloed — and they want that in a way that makes it easy for AI in a standard way to get there. That is what MCP does,” Mueller told VentureBeat. “The devil is in the details. How good is the access, how well does it perform and what does it cost. Access and governance still need to be sorted out.”
The Fabric IQ announcements arrive alongside the Database Hub, now in early access, which brings Azure SQL, Azure Cosmos DB, PostgreSQL, MySQL and SQL Server under a single management and observability layer inside Fabric. The intent is to give data operations teams one place to monitor, govern and optimize their database estate without changing how each service is deployed.
Devin Pratt, research director at IDC, said the integrated direction tracks with where the broader market is heading. IDC expects that by 2029, 60% of enterprise data platforms will unify transactional and analytical workloads.
“Microsoft’s angle is to bring more of those pieces together in one coordinated approach, while rivals are moving along similar lines from different starting points,” Pratt told VentureBeat.
For data engineers responsible for making pipelines AI-ready, the practical implication of this week’s announcements is a shift in where the hard work lives.
Connecting data sources to a platform is a solved problem. Defining what that data means in business terms, and making that definition consistently available to every agent that queries it, is not.
That shift has a concrete implication for data professionals. The semantic layer — the ontology that maps business entities, relationships and operational rules — is becoming production infrastructure. It will need to be built, versioned, governed and maintained with the same discipline as a data pipeline. That is a new category of responsibility for data engineering teams, and most organizations have not yet staffed or structured for it.
The broader trend this week’s announcements reflect is that the data platform race in 2026 is no longer primarily about compute or storage. It is about which platform can deliver the most reliable shared context to the widest range of agents.
When an AI agent loses context mid-task because traditional storage can’t keep pace with inference, it is not a model problem — it is a storage problem. At GTC 2026, Nvidia announced BlueField-4 STX, a modular reference architecture that inserts a dedicated context memory layer between GPUs and traditional storage, claiming 5x the token throughput, 4x the energy efficiency and 2x the data ingestion speed of conventional CPU-based storage.
The bottleneck STX targets is key-value cache data. KV cache is the stored record of what a model has already processed — the intermediate calculations an LLM saves so it does not have to recompute attention across the entire context on every inference step. It is what allows an agent to maintain coherent working memory across sessions, tool calls and reasoning steps. As context windows grow and agents take more steps, that cache grows with them. When it has to traverse a traditional storage path to get back to the GPU, inference slows and GPU utilization drops.
STX is not a product Nvidia sells directly. It is a reference architecture the company is distributing to its storage partner ecosystem so vendors can build AI-native infrastructure around it.
The architecture is built around a new storage-optimized BlueField-4 processor that combines Nvidia’s Vera CPU with the ConnectX-9 SuperNIC. It runs on Spectrum-X Ethernet networking and is programmable through Nvidia’s DOCA software platform.
The first rack-scale implementation is the Nvidia CMX context memory storage platform. CMX extends GPU memory with a high-performance context layer designed specifically for storing and retrieving KV cache data generated by large language models during inference. Keeping that cache accessible without forcing a round trip through general-purpose storage is what CMX is designed to do.
“Traditional data centers provide high-capacity, general-purpose storage, but generally lack the responsiveness required for interaction with AI agents that need to work across many steps, tools and different sessions,” Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing said in a briefing with press and analysts.
In response to a question from VentureBeat, Buck confirmed that STX also ships with a software reference platform alongside the hardware architecture. Nvidia is expanding DOCA to include a new component referred to in the briefing as DOCA Memo.
“Our storage providers can leverage the programmability of the BlueField-4 processor to optimize storage for the agentic AI factory,” Buck said. “In addition to having a reference rack architecture, we’re also providing a reference software platform for them to deliver those innovations and optimizations for their customers.”
Storage partners building on STX get both a hardware reference design and a software reference platform — a programmable foundation for context-optimized storage.
Storage providers co-designing STX-based infrastructure include Cloudian, DDN, Dell Technologies, Everpure, Hitachi Vantara, HPE, IBM, MinIO, NetApp, Nutanix, VAST Data and WEKA. Manufacturing partners building STX-based systems include AIC, Supermicro and Quanta Cloud Technology.
On the cloud and AI side, CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure and Vultr have all committed to STX for context memory storage.
That combination of enterprise storage incumbents and AI-native cloud providers is the signal worth watching. Nvidia is not positioning STX as a specialty product for hyperscalers. It is positioning it as the reference standard for anyone building storage infrastructure that has to serve agentic AI workloads — which, within the next two to three years, is likely to include most enterprise AI deployments running multi-step inference at scale.
STX-based platforms will be available from partners in the second half of 2026.
IBM sits on both sides of the STX announcement. It is listed as a storage provider co-designing STX-based infrastructure, and Nvidia separately confirmed that it has selected IBM Storage Scale System 6000 — certified and validated on Nvidia DGX platforms — as the high-performance storage foundation for its own GPU-native analytics infrastructure.
IBM also announced a broader expanded collaboration with Nvidia at GTC, including GPU-accelerated integration between IBM’s watsonx.data Presto SQL engine and Nvidia’s cuDF library. A production proof of concept with Nestlé put numbers on what that acceleration looks like: a data refresh cycle across the company’s Order-to-Cash data mart, covering 186 countries and 44 tables, dropped from 15 minutes to three minutes. IBM reported 83% cost savings and a 30x price-performance improvement.
The Nestlé result is a structured analytics workload. It does not directly demonstrate agentic inference performance. But it makes IBM and Nvidia’s shared argument concrete: the data layer is where enterprise AI performance is currently constrained, and GPU-accelerating it produces material results in production.
STX is a signal that the storage layer is becoming a first-class concern in enterprise AI infrastructure planning, not an afterthought to GPU procurement.
General-purpose NAS and object storage were not designed to serve KV cache data at inference latency requirements. STX-based systems from partners including Dell, HPE, NetApp and VAST Data are what Nvidia is putting forward as the practical alternative, with the DOCA software platform providing the programmability layer to tune storage behavior for specific agentic workloads.
The performance claims — 5x token throughput, 4x energy efficiency, 2x data ingestion — are measured against traditional CPU-based storage architectures. Nvidia has not specified the exact baseline configuration for those comparisons. Before those numbers drive infrastructure decisions, the baseline is worth pinning down.
Platforms are expected from partners in the second half of 2026. Given that most major storage vendors are already co-designing on STX, enterprises evaluating storage refreshes for AI infrastructure in the next 12 months should expect STX-based options to be available from their existing vendor relationships.
What’s the role of vector databases in the agentic AI world? That’s a question that organizations have been coming to terms with in recent months.
The narrative had real momentum. As large language models scaled to million-token context windows, a credible argument circulated among enterprise architects: purpose-built vector search was a stopgap, not infrastructure. Agentic memory would absorb the retrieval problem. Vector databases were a RAG-era artifact.
The production evidence is running the other way.
Qdrant, the Berlin-based open source vector search company, announced a $50 million Series B on Thursday, two years after a $28 million Series A. The timing is not incidental. The company is also shipping version 1.17 of its platform. Together, they reflect a specific argument: The retrieval problem did not shrink when agents arrived. It scaled up and got harder.
“Humans make a few queries every few minutes,” Andre Zayarni, Qdrant’s CEO and co-founder, told VentureBeat. “Agents make hundreds or even thousands of queries per second, just gathering information to be able to make decisions.”
That shift changes the infrastructure requirements in ways that RAG-era deployments were never designed to handle.
Agents operate on information they were never trained on: proprietary enterprise data, current information, millions of documents that change continuously. Context windows manage session state. They don’t provide high-recall search across that data, maintain retrieval quality as it changes, or sustain the query volumes autonomous decision-making generates.
“The majority of AI memory frameworks out there are using some kind of vector storage,” Zayarni said.
The implication is direct: even the tools positioned as memory alternatives rely on retrieval infrastructure underneath.
Three failure modes surface when that retrieval layer isn’t purpose-built for the load. At document scale, a missed result is not a latency problem — it is a quality-of-decision problem that compounds across every retrieval pass in a single agent turn. Under write load, relevance degrades because newly ingested data sits in unoptimized segments before indexing catches up, making searches over the freshest data slower and less accurate precisely when current information matters most. Across distributed infrastructure, a single slow replica pushes latency across every parallel tool call in an agent turn — a delay a human user absorbs as inconvenience but an autonomous agent cannot.
Qdrant’s 1.17 release addresses each directly. A relevance feedback query improves recall by adjusting similarity scoring on the next retrieval pass using lightweight model-generated signals, without retraining the embedding model. A delayed fan-out feature queries a second replica when the first exceeds a configurable latency threshold. A new cluster-wide telemetry API replaces node-by-node troubleshooting with a single view across the entire cluster.
Nearly every major database now supports vectors as a data type — from hyperscalers to traditional relational systems. That shift has changed the competitive question. The data type is now table stakes. What remains specialized is retrieval quality at production scale.
That distinction is why Zayarni no longer wants Qdrant called a vector database.
“We’re building an information retrieval layer for the AI age,” he said. “Databases are for storing user data. If the quality of search results matters, you need a search engine.”
His advice for teams starting out: use whatever vector support is already in your stack. The teams that migrate to purpose-built retrieval do so when scale forces the issue.
“We see companies come to us every day saying they started with Postgres and thought it was good enough — and it’s not.”
Qdrant’s architecture, written in Rust, gives it memory efficiency and low-level performance control that higher-level languages don’t match at the same cost. The open source foundation compounds that advantage — community feedback and developer adoption are what allow a company at Qdrant’s scale to compete with vendors that have far larger engineering resources.
“Without it, we wouldn’t be where we are right now at all,” Zayarni said.
The companies building production AI systems on Qdrant are making the same argument from different directions: agents need a retrieval layer, and conversational or contextual memory is not a substitute for it.
GlassDollar helps enterprises including Siemens and Mahle evaluate startups. Search is the core product: a user describes a need in natural language and gets back a ranked shortlist from a corpus of millions of companies. The architecture runs query expansion on every request – a single prompt fans out into multiple parallel queries, each retrieving candidates from a different angle, before results are combined and re-ranked. That is an agentic retrieval pattern, not a RAG pattern, and it requires purpose-built search infrastructure to sustain it at volume.
The company migrated from Elasticsearch as it scaled toward 10 million indexed documents. After moving to Qdrant it cut infrastructure costs by roughly 40%, dropped a keyword-based compensation layer it had maintained to offset Elasticsearch’s relevance gaps, and saw a 3x increase in user engagement.
“We measure success by recall,” Kamen Kanev, GlassDollar’s head of product, told VentureBeat. “If the best companies aren’t in the results, nothing else matters. The user loses trust.”
Agentic memory and extended context windows aren’t enough to absorb the workload that GlassDollar needs, either.
“That’s an infrastructure problem, not a conversation state management task,” Kanev said. “It’s not something you solve by extending a context window.”
Another Qdrant user is &AI, which is building infrastructure for patent litigation. Its AI agent, Andy, runs semantic search across hundreds of millions of documents spanning decades and multiple jurisdictions. Patent attorneys will not act on AI-generated legal text, which means every result the agent surfaces has to be grounded in a real document.
“Our whole architecture is designed to minimize hallucination risk by making retrieval the core primitive, not generation,” Herbie Turner, &AI’s founder and CTO, told VentureBeat.
For &AI, the agent layer and the retrieval layer are distinct by design.
“Andy, our patent agent, is built on top of Qdrant,” Turner said. “The agent is the interface. The vector database is the ground truth.”
The practical starting point: use whatever vector capability is already in your stack. The evaluation question isn’t whether to add vector search — it’s when your current setup stops being adequate. Three signals mark that point: retrieval quality is directly tied to business outcomes; query patterns involve expansion, multi-stage re-ranking, or parallel tool calls; or data volume crosses into the tens of millions of documents.
At that point the evaluation shifts to operational questions: how much visibility does your current setup give you into what’s happening across a distributed cluster, and how much performance headroom does it have when agent query volumes increase.
“There’s a lot of noise right now about what replaces the retrieval layer,” Kanev said. “But for anyone building a product where retrieval quality is the product, where missing a result has real business consequences, you need dedicated search infrastructure.”
Yesterday amid a flurry of enterprise AI product updates, Google announced arguably its most significant one for enterprise customers: the public preview availability of Gemini Embedding 2, its new embeddings model — a significant evolution in how machines represent and retrieve information across different media types.
While previous embedding models were largely restricted to text, this new model natively integrates text, images, video, audio, and documents into a single numerical space — reducing latency by as much as 70% for some customers and reducing total cost for enterprises who use AI models powered by their own data to complete business tasks.
VentureBeat collaborator Sam Witteveen, co-founder of AI and ML training company Red Dragon AI, received early access to Gemini Embedding 2 and published a video of his impressions on YouTube. Watch it below:
For those who have encountered the term “embeddings” in AI discussions but find it abstract, a useful analogy is that of a universal library.
In a traditional library, books are organized by metadata: author, title, or genre. In the “embedding space” of an AI, information is organized by ideas.
Imagine a library where books aren’t organized by the Dewey Decimal System, but by their “vibe” or “essence”. In this library, a biography of Steve Jobs would physically fly across the room to sit next to a technical manual for a Macintosh. A poem about a sunset would drift toward a photography book of the Pacific Coast, with all thematically similar content organized in beautiful hovering “clouds” of books. This is basically what an embedding model does.
An embedding model takes complex data—like a sentence, a photo of a sunset, or a snippet of a podcast—and converts it into a long list of numbers called a vector.
These numbers represent coordinates in a high-dimensional map. If two items are “semantically” similar (e.g., a photo of a golden retriever and the text “man’s best friend”), the model places their coordinates very close to each other in this map. Today, these models are the invisible engine behind:
Search Engines: Finding results based on what you mean, not just the specific words you typed.
Recommendation Systems: Netflix or Spotify suggesting content because its “coordinates” are near things you already like.
Enterprise AI: Large companies use them for Retrieval-Augmented Generation (RAG), where an AI assistant “looks up” a company’s internal PDFs to answer an employee’s question accurately.
The concept of mapping words to vectors dates back to the 1950s with linguists like John Rupert Firth, but the modern “vector revolution” began in the early 2000s when Yoshua Bengio’s team first used the term “word embeddings”. The real breakthrough for the industry was Word2Vec, released by a team at Google led by Tomas Mikolov in 2013. Today, the market is led by a handful of major players:
OpenAI: Known for its widely-used text-embedding-3 series.
Google: With the new Gemini and previous Gecko models.
Anthropic and Cohere: Providing specialized models for enterprise search and developer workflows.
By moving beyond text to a natively multimodal architecture, Google is attempting to create a singular, unified map for the sum of human digital expression—text, images, video, audio, and documents—all residing in the same mathematical neighborhood.
Most leading models are still “text-first.” If you want to search a video library, the AI usually has to transcribe the video into text first, then embed that text.
Google’s Gemini Embedding 2 is natively multimodal.
As Logan Kilpatrick of Google DeepMind posted on X, the model allows developers to “bring text, images, video, audio, and docs into the same embedding space”.
It understands audio as sound waves and video as motion directly, without needing to turn them into text first. This reduces “translation” errors and captures nuances that text alone might miss.
For developers and enterprises, the “natively multimodal” nature of Gemini Embedding 2 represents a shift toward more efficient AI pipelines.
By mapping all media into a single 3,072-dimensional space, developers no longer need separate systems for image search and text search; they can perform “cross-modal” retrieval—using a text query to find a specific moment in a video or an image that matches a specific sound.
And unlike its predecessors, Gemini Embedding 2 can process requests that mix modalities. A developer can send a request containing both an image of a vintage car and the text “What is the engine type?”. The model doesn’t process them separately; it treats them as a single, nuanced concept. This allows for a much deeper understanding of real-world data where the “meaning” is often found in the intersection of what we see and what we say.
One of the model’s more technical features is Matryoshka Representation Learning. Named after Russian nesting dolls, this technique allows the model to “nest” the most important information in the first few numbers of the vector.
An enterprise can choose to use the full 3072 dimensions for maximum precision, or “truncate” them down to 768 or 1536 dimensions to save on database storage costs with minimal loss in accuracy.
Gemini Embedding 2 establishes a new performance ceiling for multimodal depth, specifically outperforming previous industry leaders across text, image, and video evaluation tasks.
The model’s most significant lead is found in video and audio retrieval, where its native architecture allows it to bypass the performance degradation typically associated with text-based transcription pipelines.
Specifically, in video-to-text and text-to-video retrieval tasks, the model demonstrates a measurable performance gap over existing industry leaders, accurately mapping motion and temporal data into a unified semantic space.
The technical results show a distinct advantage in the following standardized categories:
Multimodal Retrieval: Gemini Embedding 2 consistently outperforms leading text and vision models in complex retrieval tasks that require understanding the relationship between visual elements and textual queries.
Speech and Audio Depth: The model introduces a new standard for native audio embeddings, achieving higher accuracy in capturing phonetic and tonal intent compared to models that rely on intermediate text-transcription.
Contextual Scaling: In text-based benchmarks, the model maintains high precision while utilizing its expansive 8,192 token context window, ensuring that long-form documents are embedded with the same semantic density as shorter snippets.
Dimension Flexibility: Testing across the Matryoshka Representation Learning (MRL) layers reveals that even when truncated to 768 dimensions, the model retains a significant majority of its 3,072-dimension performance, outperforming fixed-dimension models of similar size.
For the modern enterprise, information is often a fragmented mess. A single customer issue might involve a recorded support call (audio), a screenshot of an error (image), a PDF of a contract (document), and a series of emails (text).
In previous years, searching across these formats required four different pipelines. With Gemini Embedding 2, an enterprise can create a Unified Knowledge Base. This enables a more advanced form of RAG, wherein a company’s internal AI doesn’t just look up facts, but understands the relationship between them regardless of format.
Early partners are already reporting drastic efficiency gains:
Sparkonomy, a creator economy platform, reported that the model’s native multimodality slashed their latency by up to 70%. By removing the need for intermediate LLM “inference” (the step where one model explains a video to another), they nearly doubled their semantic similarity scores for matching creators with brands.
Everlaw, a legal tech firm, is using the model to navigate the “high-stakes setting” of litigation discovery. In legal cases where millions of records must be parsed, Gemini’s ability to index images and videos alongside text allows legal professionals to find “smoking gun” evidence that traditional text-search would miss.
In its announcement, Google was upfront about some of the current limitations of Gemini Embedding 2. The new model can accommodate vectorization of individual files that comprise of as many as 8,192 text tokens, 6 images (in as single batch), 128 seconds of video (2 minutes, 8 seconds long), 80 seconds of native audio (1.34 minutes), and a 6-page PDF.
It is vital to clarify that these are input limits per request, not a cap on what the system can remember or store.
Think of it like a scanner. If a scanner has a limit of “one page at a time,” it doesn’t mean you can only ever scan one page. it means you have to feed the pages in one by one.
Individual File Size: You cannot “embed” a 100-page PDF in a single call. You must “chunk” the document—splitting it into segments of 6 pages or fewer—and send each segment to the model individually.
Cumulative Knowledge: Once those chunks are converted into vectors, they can all live together in your database. You can have a database containing ten million 6-page PDFs, and the model will be able to search across all of them simultaneously.
Video and Audio: Similarly, if you have a 10-minute video, you would break it into 128-second segments to create a searchable “timeline” of embeddings.
As of March 10, 2026, Gemini Embedding 2 is officially in Public Preview.
For developers and enterprise leaders, this means the model is accessible for immediate testing and production integration, though it is still subject to the iterative refinements typical of “preview” software before it reaches General Availability (GA).
The model is deployed across Google’s two primary AI gateways, each catering to a different scale of operation:
Gemini API: Targeted at rapid prototyping and individual developers, this path offers a simplified pricing structure.
Vertex AI (Google Cloud): The enterprise-grade environment designed for massive scale, offering advanced security controls and integration with the broader Google Cloud ecosystem.
It’s also already integrated with the heavy hitters of AI infrastructure: LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, and ChromaDB.
In the Gemini API, Google has introduced a tiered pricing model that distinguishes between “standard” data (text, images, and video) and “native” audio.
The Free Tier: Developers can experiment with the model at no cost, though this tier comes with rate limits (typically 60 requests per minute) and uses data to improve Google’s products.
The Paid Tier: For production-level volume, the cost is calculated per million tokens. For text, image, and video inputs, the rate is $0.25 per 1 million tokens.
The “Audio Premium”: Because the model natively ingests audio data without intermediate transcription—a more computationally intensive task—the rate for audio inputs is doubled to $0.50 per 1 million tokens.
For large-scale deployments on Vertex AI, the pricing follows an enterprise-centric “Pay-as-you-go” (PayGo) model. This allows organizations to pay for exactly what they use across different processing modes:
Flex PayGo: Best for unpredictable, bursty workloads.
Provisioned Throughput: Designed for enterprises that require guaranteed capacity and consistent latency for high-traffic applications.
Batch Prediction: Ideal for re-indexing massive historical archives, where time-sensitivity is lower but volume is extremely high.
By making the model available through these diverse channels and integrating it natively with libraries like LangChain, LlamaIndex, and Weaviate, Google has ensured that the “switching cost” for businesses isn’t just a matter of price, but of operational ease. Whether a startup is building its first RAG-based assistant or a multinational is unifying decades of disparate media archives, the infrastructure is now live and globally accessible.
In addition, the official Gemini API and Vertex AI Colab notebooks, which contain the Python code necessary to implement these features, are licensed under the Apache License, Version 2.0.
The Apache 2.0 license is highly regarded in the tech community because it is “permissive.” It allows developers to take Google’s implementation code, modify it, and use it in their own commercial products without having to pay royalties or “open source” their own proprietary code in return.
For Chief Data Officers and technical leads, the decision to migrate to Gemini Embedding 2 hinges on the transition from a “text-plus” strategy to a “natively multimodal” one.
If your organization currently relies on fragmented pipelines — where images and videos are first transcribed or tagged by separate models before being indexed — the upgrade is likely a strategic necessity.
This model eliminates the “translation tax” of using intermediate LLMs to describe visual or auditory data, a move that partners like Sparkonomy found reduced latency by up to 70% while doubling semantic similarity scores. For businesses managing massive, diverse datasets, this isn’t just a performance boost; it is a structural simplification that reduces the number of points where “meaning” can be lost or distorted.
The effort to switch from a text-only foundation is lower than one might expect due to what early users describe as excellent “API continuity”.
Because the model integrates with industry-standard frameworks like LangChain, LlamaIndex, and Vector Search, it can often be “dropped into” existing workflows with minimal code changes. However, the real cost and energy investment lies in re-indexing. Moving to this model requires re-embedding your existing corpus to ensure all data points exist in the same 3,072-dimensional space.
While this is a one-time computational hurdle, it is the prerequisite for unlocking cross-modal search—where a simple text query can suddenly “see” into your video archives or “hear” specific customer sentiment in call recordings.
The primary trade-off for data leaders to weigh is the balance between high-fidelity retrieval and long-term storage economics. Gemini Embedding 2 addresses this directly through Matryoshka Representation Learning (MRL), which allows you to truncate vectors from 3072 dimensions down to 768 without a linear drop in quality.
This gives CDOs a tactical lever: you can choose maximum precision for high-stakes legal or medical discovery—as seen in Everlaw’s 20% lift in recall—while utilizing smaller, more efficient vectors for lower-priority recommendation engines to keep cloud storage costs in check.
Ultimately, the ROI is found in the “lift” of accuracy; in a landscape where an AI’s value is defined by its context, the ability to natively index a 6-page PDF or 128 seconds of video directly into a knowledge base provides a depth of insight that text-only models simply cannot replicate.
OpenAI on Monday launched a set of interactive visual tools inside ChatGPT that let users manipulate mathematical and scientific formulas in real time — a genuinely impressive education feature that also serves as the company’s most direct attempt yet to change the subject during the worst ten days of its corporate life.
The new experience covers more than 70 core math and science concepts, from the Pythagorean theorem to Ohm’s law to compound interest. When a user asks ChatGPT to explain one of these topics, the chatbot now generates a dynamic module with adjustable sliders alongside its written response. Drag a variable, and the equations, graphs, and diagrams update instantly. The feature is available today to all logged-in users worldwide, across every plan, including free.
OpenAI tells VentureBeat that 140 million people already use ChatGPT each week for math and science learning. That is a staggering number — and it goes a long way toward explaining why the company chose this particular week to ship a product designed to make those users’ experience meaningfully better. Since late February, OpenAI has been sued by the family of a 12-year-old mass shooting victim who alleges the company knew the attacker was planning violence through ChatGPT; lost its head of robotics over a Pentagon deal that triggered a near-300% spike in app uninstalls; watched more than 30 of its own employees file a legal brief supporting rival Anthropic against the U.S. government; and scrapped plans with Oracle to expand a flagship data center in Texas. Its chief competitor’s app, Claude, now sits atop the App Store.
The interactive learning tools are, on their merits, a strong product. But they arrive at a company fighting on every front simultaneously — and burning through an estimated $15 billion in cash this year to do it.
The feature is built on a simple pedagogical premise: students understand formulas better when they can see what happens as the inputs change.
Ask ChatGPT “help me understand the Pythagorean theorem,” and the system now responds with a written explanation alongside an interactive panel. On the left, the formula $a^2 + b^2 = c^2$ appears in clean notation with sliders for sides $a$ and $b$. On the right, a geometric visualization — a right triangle with squares drawn on each side — reshapes dynamically as you adjust the values. The computed hypotenuse updates in real time. The same treatment applies across topics: voltage and resistance for Ohm’s law, pressure and temperature for the ideal gas equation, radius and height for cone volume.
OpenAI’s initial roster of more than 70 topics targets high school and introductory college material: binomial squares, Charles’ law, circle equations, Coulomb’s law, cylinder volume, degrees of freedom, exponential decay, Hooke’s law, kinetic energy, the lens equation, linear equations, slope-intercept form, surface area of a sphere, trigonometric angle sum identities, and others.
The company cited research suggesting that “visual, interaction-based learning can lead to stronger conceptual understanding than traditional instruction for many students,” and pointed to a recent Gallup survey in which more than half of U.S. adults said they struggle with math. In early testing, OpenAI said, students reported the modules helped them grasp how variables relate to one another, and parents described using them to work through problems alongside their children.
Anjini Grover, a high school mathematics teacher quoted in OpenAI’s announcement, said the feature stands out for “how strongly this feature emphasizes conceptual understanding.” Raquel Gibson, a high school algebra teacher, called it “a step towards empowering students to independently explore abstract concepts.”
The tools build on ChatGPT’s existing education features — a “study mode” for step-by-step problem solving and a quizzes feature for exam prep — and OpenAI said it plans to expand interactive learning to additional subjects. The company also said it intends to publish research through its NextGenAI initiative and OpenAI Learning Lab to study how AI shapes learning outcomes over time.
The education launch shares the calendar with the most serious legal challenge OpenAI has ever faced.
On Monday, the mother of 12-year-old Maya Gebala filed a civil lawsuit against OpenAI in B.C. Supreme Court, alleging the company had “specific knowledge of the shooter’s long-range planning of a mass casualty event” through ChatGPT interactions and “took no steps to act upon this knowledge.” Gebala was shot three times during a mass shooting in Tumbler Ridge, British Columbia on February 10 that killed eight people and the 18-year-old attacker. She suffered what the lawsuit describes as a catastrophic traumatic brain injury with permanent cognitive and physical disabilities.
The claim paints a damning picture of how the shooter used ChatGPT. It alleges the platform functioned as a “counsellor, pseudo-therapist, trusted confidante, friend, and ally” and was “intentionally designed to foster psychological dependency between the user and ChatGPT.” The shooter was under 18 when they began using the service, the suit states, and despite OpenAI’s requirement that minors obtain parental consent, the company “took no steps to implement age verification or consent procedures.”
OpenAI has separately acknowledged that it suspended the shooter’s account months before the attack but did not alert Canadian law enforcement — a decision that provoked sharp political fallout. B.C. Premier David Eby said after a virtual meeting with Altman that the CEO agreed to apologize to the people of Tumbler Ridge and work with the provincial government on AI regulation recommendations.
None of the claims have been proven in court. OpenAI has not publicly commented on the lawsuit. But the case poses a question that transcends any single legal proceeding: when an AI company’s own internal systems identify a user as dangerous enough to ban, what obligation does it have to tell someone?
The Tumbler Ridge lawsuit is unfolding against the backdrop of an internal crisis that has already cost OpenAI key talent and millions of users.
On February 28, CEO Sam Altman announced a deal giving the Pentagon access to OpenAI’s AI models inside secure government computing systems. The agreement came days after Anthropic CEO Dario Amodei publicly refused similar terms, saying his company could not proceed without assurances against autonomous weapons and mass domestic surveillance. The Pentagon responded by designating Anthropic a “supply-chain risk” — a classification normally reserved for foreign adversaries — and Defense Secretary Pete Hegseth barred any military contractor from conducting commercial activity with the company.
The reaction inside OpenAI was immediate. Caitlin Kalinowski, who joined from Meta in 2024 to build out the company’s robotics hardware division, resigned on principle. “AI has an important role in national security,” she wrote publicly. “But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got.” Research scientist Aidan McLaughlin wrote on social media that he “personally don’t think this deal was worth it.” Another employee told CNN that many OpenAI staffers “really respect” Anthropic for walking away.
The reaction outside the company was even more dramatic. ChatGPT uninstalls spiked more than 295% on the day the deal was announced. Anthropic’s Claude surged to No. 1 among free apps on the U.S. Apple App Store and remained there as of this past weekend. Protesters gathered outside OpenAI’s San Francisco headquarters calling for a “QuitGPT” movement.
And in the most extraordinary development, more than 30 OpenAI and Google DeepMind employees — including DeepMind chief scientist Jeff Dean — filed an amicus brief Monday supporting Anthropic’s lawsuit against the Defense Department. The brief argued that the Pentagon’s actions, “if allowed to proceed,” would “undoubtedly have consequences for the United States’ industrial and scientific competitiveness in the field of artificial intelligence and beyond.” The employees signed in their personal capacity, but the spectacle of OpenAI’s own researchers rallying to a competitor’s legal defense against the same government their company just partnered with has no real precedent in the industry.
Altman, to his credit, has not pretended the situation is fine. In an internal memo later shared publicly, he admitted the deal “was definitely rushed” and “just looked opportunistic and sloppy.” He revised the contract to include explicit prohibitions against mass domestic surveillance and the use of OpenAI technology on commercially acquired data. He also publicly said that enforcing the supply-chain risk designation against Anthropic “would be very bad for our industry and our country.”
Meanwhile, Anthropic warned in court filings that the Pentagon’s blacklisting could cost it up to $5 billion in lost business — roughly equivalent to its total revenue since commercializing its AI technology in 2023. The company is seeking a temporary court order to continue working with military contractors while the case proceeds.
Strip away the lawsuits and the politics, and OpenAI still has a math problem of its own.
The company is expected to burn through approximately $15 billion in cash this year, up from $9 billion in 2025. It has roughly 910 million weekly users. About 95% of them pay nothing. Subscriptions alone cannot bridge that gap, which is why OpenAI is simultaneously building out an internal advertising infrastructure and leaning on partners like Criteo — and reportedly The Trade Desk — to bring advertisers into ChatGPT.
The company is hiring aggressively for this effort: a monetization infrastructure engineer, an engineering manager, a product designer for the ads experience, a senior manager for ad revenue accounting, and a trust and safety specialist dedicated to the ads product, all based at headquarters in San Francisco. The compensation bands run as high as $385,000 — the kind of investment a company makes when it plans to own its ad stack, not rent it.
But advertising inside ChatGPT introduces a trust problem that compounds the ones OpenAI is already managing. Users who abandoned the app over the Pentagon deal demonstrated that loyalty to ChatGPT is thinner than its market share suggests. Adding commercial messages to a product already under fire for its military ties and its handling of a mass shooter’s data will require OpenAI to navigate user sentiment with a precision it has not recently demonstrated.
The infrastructure picture is equally unsettled. Oracle and OpenAI recently scrapped plans to expand a flagship AI data center in Abilene, Texas, after negotiations stalled over financing and OpenAI’s evolving needs. Meta and Nvidia moved quickly to explore the site — a reminder that in the current AI arms race, any gap in execution gets filled by a competitor within days.
This is where the education feature becomes more than a product announcement.
Education has always been ChatGPT’s cleanest use case — the application where the technology most obviously augments human capability rather than surveilling it, weaponizing it, or monetizing the attention of people who came looking for help. It is the use case that resonates across demographics: students prepping for the SAT, parents revisiting algebra at the kitchen table, adults circling back to concepts they never quite understood. And it is the use case where ChatGPT still holds a clear lead. Google’s Gemini, Anthropic’s Claude, and xAI’s Grok are all investing in education, but none has shipped anything comparable to real-time interactive formula visualization embedded in a conversational interface.
OpenAI acknowledged that the “research landscape on how AI affects learning is still taking shape,” but pointed to its own early findings on study mode as showing “promising early signals.” The company said it will continue working with educators and researchers through its NextGenAI initiative and OpenAI Learning Lab, and plans to publish findings and expand into additional subjects.
Somewhere tonight, a ninth-grader will open ChatGPT, drag a slider, and watch a hypotenuse lengthen across her screen. The Pythagorean theorem will make sense for the first time. She will not know about the Pentagon deal, or the Tumbler Ridge lawsuit, or the 295% spike in uninstalls, or the $15 billion cash burn underwriting the server that just rendered her triangle. She will only know that it worked. For OpenAI, that may have to be enough — for now.
Anthropic on Monday released Code Review, a multi-agent code review system built into Claude Code that dispatches teams of AI agents to scrutinize every pull request for bugs that human reviewers routinely miss. The feature, now available in research preview for Team and Enterprise customers, arrives on what may be the most consequential day in the company’s history: Anthropic simultaneously filed lawsuits against the Trump administration over a Pentagon blacklisting, while Microsoft announced a new partnership embedding Claude into its Microsoft 365 Copilot platform.
The convergence of a major product launch, a federal legal battle, and a landmark distribution deal with the world’s largest software company captures the extraordinary tension defining Anthropic’s current moment. The San Francisco-based AI lab is simultaneously trying to grow a developer tools business approaching $2.5 billion in annualized revenue, defend itself against an unprecedented government designation as a national security threat, and expand its commercial footprint through the very cloud platforms now navigating the fallout.
Code Review is Anthropic’s most aggressive bet yet that engineering organizations will pay significantly more — $15 to $25 per review — for AI-assisted code quality assurance that prioritizes thoroughness over speed. It also signals a broader strategic pivot: the company isn’t just building models, it’s building opinionated developer workflows around them.
Code Review works differently from the lightweight code review tools most developers are accustomed to. When a developer opens a pull request, the system dispatches multiple AI agents that operate in parallel. These agents independently search for bugs, then cross-verify each other’s findings to filter out false positives, and finally rank the remaining issues by severity. The output appears as a single overview comment on the PR along with inline annotations for specific bugs.
Anthropic designed the system to scale dynamically with the complexity of the change. Large or intricate pull requests receive more agents and deeper analysis; trivial changes get a lighter pass. The company says the average review takes approximately 20 minutes — far slower than the near-instant feedback of tools like GitHub Copilot’s built-in review, but deliberately so.
“We built Code Review based on customer and internal feedback,” an Anthropic spokesperson told VentureBeat. “In our testing, we’ve found it provides high-value feedback and has helped catch bugs that we may have missed otherwise. Developers and engineering teams use a range of tools, and we build for that reality. The goal is to give teams a capable option at every stage of the development process.”
The system emerged from Anthropic’s own engineering practices, where the company says code output per engineer has grown 200% over the past year. That surge in AI-assisted code generation created a review bottleneck that the company says it now hears about from customers on a weekly basis. Before Code Review, only 16% of Anthropic’s internal PRs received substantive review comments. That figure has jumped to 54%.
Crucially, Code Review does not approve pull requests. That decision remains with human reviewers. Instead, the system functions as a force multiplier, surfacing issues so that human reviewers can focus on architectural decisions and higher-order concerns rather than line-by-line bug hunting.
The pricing will draw immediate scrutiny. At $15 to $25 per review, billed on token usage and scaling with PR size, Code Review is substantially more expensive than alternatives. GitHub Copilot offers code review natively as part of its existing subscription, and startups like CodeRabbit operate at significantly lower price points. Anthropic’s more basic code review GitHub Action — which remains open source — is itself a lighter-weight and cheaper option.
Anthropic frames the cost not as a productivity expense but as an insurance product. “For teams shipping to production, the cost of a shipped bug dwarfs $20/review,” the company’s spokesperson told VentureBeat. “A single production incident — a rollback, a hotfix, an on-call page — can cost more in engineer hours than a month of Code Review. Code Review is an insurance product for code quality, not a productivity tool for churning through PRs faster.”
That framing is deliberate and revealing. Rather than competing on speed or price — the dimensions where lightweight tools have an advantage — Anthropic is positioning Code Review as a depth-first tool aimed at engineering leaders who manage production risk. The implicit argument is that the real cost comparison isn’t Code Review versus CodeRabbit, but Code Review versus the fully loaded cost of a production outage, including engineer time, customer impact, and reputational damage.
Whether that argument holds up will depend on the data. Anthropic has not yet published external benchmarks comparing Code Review’s bug-detection rates against competitors, and the spokesperson did not provide specific figures on bugs caught per dollar or developer hours saved when asked directly. For engineering leaders evaluating the tool, that gap in publicly available comparative data may slow adoption, even if the theoretical ROI case is compelling.
Anthropic’s internal usage data provides an early window into the system’s performance characteristics. On large pull requests exceeding 1,000 lines changed, 84% receive findings, averaging 7.5 issues per review. On small PRs under 50 lines, that drops to 31% with an average of 0.5 issues. The company reports that less than 1% of findings are marked incorrect by engineers.
That sub-1% figure is the kind of stat that demands careful unpacking. When asked how “marked incorrect” is defined, the Anthropic spokesperson explained that it means “an engineer actively resolving the comment without fixing it. We’ll continue to monitor feedback and engagement while Code Review is in research preview.”
The methodology matters. This is an opt-in disagreement metric — an engineer has to take the affirmative step of dismissing a finding. In practice, developers under time pressure may simply ignore irrelevant findings rather than actively marking them as wrong, which would cause false positives to go uncounted. Anthropic acknowledged the limitation implicitly by noting the system is in research preview and that it will continue monitoring engagement data. The company has not yet conducted or published a controlled evaluation comparing agent findings against a ground-truth baseline established by expert human reviewers.
The anecdotal evidence is nonetheless striking. Anthropic described a case where a one-line change to a production service — the kind of diff that typically receives a cursory approval — was flagged as critical by Code Review because it would have broken authentication for the service. In another example involving TrueNAS’s open-source middleware, Code Review surfaced a pre-existing bug in adjacent code during a ZFS encryption refactor: a type mismatch that was silently wiping the encryption key cache on every sync. These are precisely the categories of bugs — latent issues in touched-but-unchanged code, and subtle behavioral changes hiding in small diffs — that human reviewers are statistically most likely to miss.
The Code Review launch does not exist in a vacuum. On the same day, Anthropic filed two lawsuits — one in the U.S. District Court for the Northern District of California and another in the D.C. Circuit Court of Appeals — challenging the Trump administration’s decision to label the company a supply chain risk to national security, a designation historically reserved for foreign adversaries.
The legal confrontation stems from a breakdown in contract negotiations between Anthropic and the Pentagon. As CNN reported, the Defense Department wanted unrestricted access to Claude for “all lawful purposes,” while Anthropic insisted on two redlines: that its AI would not be used for fully autonomous weapons or mass domestic surveillance. When talks collapsed by a Pentagon-set deadline on February 27, President Trump directed all federal agencies to cease using Anthropic’s technology, and Defense Secretary Pete Hegseth formally designated the company a supply chain risk.
According to CNBC, the complaint alleges that these actions are “unprecedented and unlawful” and are “harming Anthropic irreparably,” with the company stating that contracts are already being cancelled and “hundreds of millions of dollars” in near-term revenue are in jeopardy.
“Seeking judicial review does not change our longstanding commitment to harnessing AI to protect our national security,” the Anthropic spokesperson told VentureBeat, “but this is a necessary step to protect our business, our customers, and our partners. We will continue to pursue every path toward resolution, including dialogue with the government.”
For enterprise buyers evaluating Code Review and other Claude-based tools, the lawsuit introduces a novel category of vendor risk. The supply chain risk designation doesn’t just affect Anthropic’s government contracts — as CNBC reported, it requires defense contractors to certify they don’t use Claude in their Pentagon-related work. That creates a chilling effect that could extend well beyond the defense sector, even as the company’s commercial momentum accelerates.
The market’s response to the Pentagon crisis has been notably bifurcated. While the government moved to isolate Anthropic, the company’s three largest cloud distribution partners moved in the opposite direction.
Microsoft on Monday announced it is integrating Claude into Microsoft 365 Copilot through a new product called Copilot Cowork, developed in close collaboration with Anthropic. As Yahoo Finance reported, the service enables enterprise users to perform tasks like building presentations, pulling data into Excel spreadsheets, and coordinating meetings — the kind of agentic productivity capabilities that sent shares of SaaS companies like Salesforce, ServiceNow, and Intuit tumbling when Anthropic first debuted its Cowork product on January 30.
The timing is not coincidental. As TechCrunch reported last week, Microsoft, Google, and Amazon Web Services all confirmed that Claude remains available to their customers for non-defense workloads. Microsoft’s legal team specifically concluded that “Anthropic products, including Claude, can remain available to our customers — other than the Department of War — through platforms such as M365, GitHub, and Microsoft’s AI Foundry.”
That three of the world’s most powerful technology companies publicly reaffirmed their commitment to distributing Anthropic’s models — on the same day the company sued the federal government — tells enterprise customers something important about the market’s assessment of both Claude’s technical value and the legal durability of the supply chain risk designation.
For organizations considering Code Review, the data handling question looms especially large. The system necessarily ingests proprietary source code to perform its analysis. Anthropic’s spokesperson addressed this directly: “Anthropic does not train models on our customers’ data. This is part of why customers in highly regulated industries, from Novo Nordisk to Intuit, trust us to deploy AI safely and effectively.”
The spokesperson did not detail specific retention policies or compliance certifications when asked, though the company’s reference to pharmaceutical and financial services clients suggests it has undergone the kind of security review those industries require.
Administrators get several controls for managing costs and scope, including monthly organization-wide spending caps, repository-level enablement, and an analytics dashboard tracking PRs reviewed, acceptance rates, and total costs. Once enabled, reviews run automatically on new pull requests with no per-developer configuration required.
The revenue figure Anthropic confirmed — a $2.5 billion run rate as of February 12 for Claude Code — underscores just how quickly developer tooling has become a material revenue line for the company. The spokesperson pointed to Anthropic’s recent Series G fundraise for additional context but did not break out what share of total company revenue Claude Code now represents.
Code Review is available now in research preview for Claude Code Team and Enterprise plans. Whether it can justify its premium in a market already crowded with cheaper alternatives will depend on whether Anthropic can convert anecdotal bug catches and internal usage stats into the kind of rigorous, externally validated evidence that engineering leaders with production budgets require — all while navigating a legal and political environment unlike anything the AI industry has previously faced.