Cerebras Breaks Nvidia's $10B Grip: OpenAI's Massive Chip Gambit

So, What Actually Happened?

We scanned 190,000 articles this week so you don't have to, and the AI chip wars just entered a new phase. Cerebras just landed a deal with OpenAI worth over $10 billion—the largest AI infrastructure deal in history that doesn't involve Nvidia. Meanwhile, Anthropic launched Claude for Healthcare to automate administrative workflows, and New York's RAISE Act is creating the most comprehensive AI safety framework in the US. And a sobering reality check: 65% of organizations are struggling to achieve AI success according to a new DDN report.

The Bottom Line: The AI infrastructure landscape is fragmenting in real-time—and the winners won't be the ones with the most compute, but the ones who figure out how to actually use it.

The Tracks That Matter

1. Cerebras Lands $10B+ OpenAI Deal, Breaking Nvidia's Grip

Cerebras scores OpenAI deal worth over $10 billion.

This is the most significant AI infrastructure deal since the GPU shortage began. Cerebras—the company that builds wafer-scale chips the size of dinner plates—just secured a contract with OpenAI worth over $10 billion. That's not a typo. Ten billion dollars, going to a company that isn't Nvidia.

The strategic implications are enormous. OpenAI has been Nvidia's largest customer, and this deal signals that even the biggest AI labs are actively diversifying their chip supply. Cerebras' architecture is fundamentally different from GPUs—it's designed from the ground up for AI workloads, with a single massive chip instead of thousands of smaller ones working in parallel.

For the AI chip ecosystem, this is the crack in the dam. If OpenAI is willing to bet billions on non-Nvidia silicon, every other AI company will be asking why they haven't done the same. The chip monopoly narrative just got a lot more complicated.

Here's what works: If you're locked into GPU-based infrastructure, start evaluating alternatives now. The Cerebras deal suggests the economics of non-GPU AI compute are becoming competitive at scale. Your procurement team should be watching this space.

2. Anthropic Brings Claude to Healthcare

Anthropic Launches Claude for Healthcare to Automate Administrative Workflows.

Anthropic just made its biggest vertical bet yet, launching Claude for Healthcare with tools specifically designed for health system administrative work. The focus isn't on diagnosis or clinical decision-making—it's on the mountains of paperwork that drain clinician time: prior authorizations, documentation, data analysis.

The approach is strategically smart. Healthcare AI has been plagued by overpromise and underdelivery, with startups claiming to revolutionize diagnosis while hospitals struggle to get basic scheduling to work. Anthropic is going after the unglamorous but enormous problem of administrative burden, which consumes an estimated 30% of healthcare spending.

The HIPAA implications are significant. Healthcare data can't just flow to any cloud AI service, and Anthropic's positioning suggests they've built the compliance infrastructure to handle protected health information properly. That's a competitive moat that takes years to build.

Here's what works: If you're in healthcare IT, evaluate Claude for Healthcare against your administrative workflow bottlenecks. The prior authorization use case alone could justify the integration effort—it's one of the most consistently broken processes in American healthcare.

3. New York's RAISE Act Creates Comprehensive AI Safety Framework

New York Laws ”RAISE” The Bar In Addressing AI Safety.

New York just passed the RAISE Act (Responsible AI Safety in Education), creating the most comprehensive state-level framework for AI safety in the US. The law specifically targets AI companion models—those designed for ongoing personal interaction—with requirements for content filtering, parental controls, and transparency about AI capabilities and limitations.

The implications extend beyond education. New York's approach establishes precedent for how AI companions should be regulated: clear disclosure that users are interacting with AI, mandatory safety features, and accountability mechanisms. Other states are watching, and the pattern from New York's previous consumer protection laws suggests these requirements will spread.

For AI developers building consumer-facing products, this is a preview of the regulatory environment coming nationwide. The days of deploying AI companions without safety infrastructure are ending. The question isn't whether to build these capabilities, but how quickly you can get them in place.

Here's what works: Review your AI products against the RAISE Act requirements, even if you don't operate in New York. These will likely become the baseline for US AI companion regulation within 18 months.

4. Databricks Genie Powers Atlassian Rovo's Conversational Insights

Databricks Genie Powers Conversational Insights in Atlassian Rovo.

Databricks' Genie—their natural language interface for querying data—is now powering conversational analytics inside Atlassian Rovo. This isn't a vague partnership announcement; it's a production integration that lets users ask questions in plain English and get answers drawn from their company's data lakehouse.

The architecture is significant. Rather than moving data to a separate AI system, Genie queries data where it lives in Databricks, respecting existing access controls and governance policies. For enterprises with petabytes of data and complex permission structures, this solves a real problem: how do you democratize data access without creating security nightmares?

The Atlassian integration matters because Rovo is becoming the AI layer across Jira, Confluence, and Atlassian's entire product suite. If you're already using Atlassian tools and Databricks for analytics, this integration collapses two separate workflows into one.

Here's what works: If you're running both Databricks and Atlassian, evaluate this integration. The ability to query your lakehouse from inside the tools your team already uses could significantly reduce the friction between ”having data” and ”using data.”

5. France Fines Telcos €42M for Data Breach

France fines telcos €42M for issues leading to 2024 breach.

France's data protection authority just hit Free and Free Mobile with a combined €42 million fine for security failures that led to their 2024 breach affecting over 24 million customers. The fine targets specific security failures: inadequate authentication mechanisms, insufficient encryption of stored data, and failure to detect the breach in a timely manner.

The enforcement action is notable for its specificity. CNIL isn't just saying ”you had a breach”—they're detailing exactly which security controls were inadequate and why. For other organizations, this creates a checklist: are your authentication systems robust? Is your data encrypted at rest? Do you have detection mechanisms that would catch this kind of exfiltration?

The €42 million figure also signals that European regulators are getting serious about breach accountability. This isn't a slap on the wrist—it's a material financial penalty that affects the bottom line.

Here's what works: Use the CNIL findings as an internal audit checklist. The specific security failures they identified—authentication weaknesses, encryption gaps, detection failures—are common across industries. Better to find them yourself than have a regulator do it for you.

6. DDN Report: 65% of Organizations Struggling to Achieve AI Success

DDN Report: 65% of Organizations Struggling to Achieve AI Success.

A new report from DDN—the data infrastructure company—reveals that 65% of organizations are struggling to achieve AI success, despite massive investments in AI capabilities. The primary culprits: data quality issues, infrastructure bottlenecks, and the gap between AI experimentation and production deployment.

The findings cut against the dominant narrative that AI adoption is straightforward. The report suggests that most organizations have successfully run AI proofs of concept, but far fewer have managed to deploy AI in production at scale. The bottleneck isn't AI capability—it's everything around it: data pipelines, infrastructure, integration, and operational support.

For enterprise leaders, this is validation that AI struggles aren't unique to their organization. The ”everyone else is doing AI successfully” fear is largely unfounded. Most organizations are struggling with the same challenges: getting clean data to models, maintaining model performance over time, and integrating AI outputs into existing workflows.

Here's what works: Shift your AI metrics from ”did we build a model?” to ”is it running in production, improving outcomes?” The gap between POC and production is where most AI initiatives die.

7. Defense Tech Onebrief Valued at Over $2 Billion

Defense tech firm Onebrief valued at over $2 billion in latest funding round.

Onebrief, the defense planning software company, just hit a $2 billion+ valuation in its latest funding round. The company builds AI-powered tools for military planning and operations, automating workflows that traditionally required extensive staff time.

The valuation reflects a broader trend: defense technology is attracting venture capital at rates not seen since the early 2000s. The combination of AI capabilities, geopolitical uncertainty, and modernization mandates has created a funding environment where defense-focused startups can command enterprise software multiples.

For commercial enterprise software companies, the defense sector offers lessons. Onebrief's success is built on workflow automation and planning tools—capabilities that translate directly to commercial applications. The difference is that defense procurement, despite its reputation for slowness, is currently moving faster than many commercial enterprises on AI adoption.

Here's what works: Watch defense tech innovations for commercial applications. The planning and automation tools being built for military use cases often translate well to enterprise operations, and defense-funded R&D tends to produce robust, production-ready systems.

Signal vs. Noise

🟢 Signal: Data breach accountability is having its moment. The €42M French fine, combined with rising insurance costs and regulatory scrutiny, suggests that breach response is evolving from ”apologize and move on” to ”pay real consequences.” Organizations that invest in security infrastructure now will avoid both the breach costs and the regulatory penalties.

🔴 Noise: ”AI in Healthcare” headlines continue to outpace actual deployments. While Anthropic's Claude for Healthcare is a real product with real capabilities, the broader healthcare AI narrative remains more promise than delivery. Be skeptical of healthcare AI announcements that don't include specific workflow integrations and customer deployments.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

The Infrastructure Diversification Accelerates

Three developments this week tell a connected story: Cerebras landing the $10B+ OpenAI deal, Databricks deepening its enterprise integrations, and DDN reporting that 65% of organizations are struggling with AI success.

The common thread: AI infrastructure is fragmenting, and the organizations that succeed will be those that pick the right pieces for their specific needs. The one-size-fits-all approach—buy GPUs, run OpenAI models, figure out the rest—is breaking down. Cerebras is proving that non-GPU compute can work at the highest scale. Databricks is showing that data infrastructure and AI need to be integrated, not separate. DDN is documenting that infrastructure choices matter more than model capabilities for production success.

The organizations stuck at 65% failure aren't failing because they picked the wrong AI model. They're failing because they treated infrastructure as an afterthought. The Cerebras deal suggests that even OpenAI—the most sophisticated AI company in the world—is rethinking its infrastructure assumptions.

The implication: Your AI strategy needs an infrastructure strategy. The chip, the data platform, the integration layer—these aren't implementation details. They're strategic decisions that will determine whether your AI investments deliver.

By The Numbers

• $10B+ — Cerebras' deal with OpenAI, the largest non-Nvidia AI infrastructure contract
• 65% — Organizations struggling to achieve AI success according to DDN report
• €42M — France's fine against Free and Free Mobile for 2024 data breach
• $2B+ — Onebrief's valuation for defense planning AI
• $80M — Proxima's raise for AI-powered drug discovery
• 30% — Estimated healthcare spending consumed by administrative burden

Deep Dive: The 65% Problem

Like a DJ who realizes most of the crowd isn't dancing, the AI industry is confronting an uncomfortable truth: the majority of organizations aren't succeeding with AI, despite years of investment and hype.

What the Data Shows

DDN's report puts hard numbers on what many enterprise leaders have suspected: 65% of organizations are struggling to achieve AI success. Not ”haven't started”—struggling. These are companies that have bought the GPUs, hired the data scientists, and built the models. They're stuck somewhere between proof of concept and production value.

The failure modes are consistent: data quality issues prevent models from learning what they need to learn. Infrastructure bottlenecks create latency that makes real-time AI impossible. Integration challenges mean AI outputs don't reach the workflows where they'd create value. And operational support—monitoring, retraining, error handling—doesn't exist at the scale production requires.

Why Infrastructure Matters More Than Models

The Cerebras-OpenAI deal illuminates a key insight: even the most sophisticated AI company in the world is prioritizing infrastructure diversification. OpenAI has access to every model architecture, every training technique, every AI capability. What they're investing $10 billion in is alternative compute infrastructure.

The implication for enterprise AI is significant. The limiting factor for most organizations isn't model capability—it's everything else. The best model in the world is useless if your data pipeline feeds it garbage, your infrastructure can't run it at acceptable latency, and your workflows don't consume its outputs.

What Actually Works

1. Audit your data quality ruthlessly: The most common AI failure mode is training on bad data. Before investing in more capable models, ensure your data pipeline produces data worth training on.

2. Match infrastructure to workload: The Cerebras deal shows that GPU dominance isn't guaranteed. Evaluate whether your specific AI workloads might benefit from alternative architectures—particularly for inference at scale.

3. Focus on integration, not models: The organizations succeeding with AI have invested heavily in connecting AI outputs to existing workflows. The model is 20% of the work; integration is 80%.

4. Build operational capabilities: Production AI requires monitoring, retraining pipelines, and error handling. If your team is still manually retraining models, you don't have production AI—you have ongoing research.

5. Set realistic timelines: The 65% struggling aren't failing because AI doesn't work. They're failing because they expected immediate results from investments that take 2-3 years to mature.

The 65% problem isn't a condemnation of AI. It's a correction in expectations. AI creates enormous value—for the organizations that solve the infrastructure, integration, and operational challenges that make AI deployable.

What's Coming

DeepSeek's Engram Technique Advances V4 Model

DeepSeek Proposes Workaround for AI Training Before V4 Model. The Chinese AI lab is developing novel training techniques that could reduce compute requirements for large models. Watch for DeepSeek to continue challenging Western AI labs on efficiency.

Sovereign AI and Data Residency Become Competitive

Sovereign AI: Data Residency as a Competitive Edge. As AI regulation proliferates, the ability to guarantee data stays within specific jurisdictions is becoming a product feature, not just a compliance requirement.

Slackbot Gets an Agentic AI Makeover

Slackbot's agentic AI makeover gives users their copilot. Salesforce is positioning Slack as an AI agent platform, with the humble Slackbot evolving into a productivity copilot. Expect more collaboration tools to follow this pattern.

For Your Team

Thursday's meeting prompt: ”DDN reports that 65% of organizations are struggling to achieve AI success. Are we in the 35% that's succeeding, or the 65% that's struggling? What's our evidence either way?”

The Infrastructure-First Framework:

1. Audit your data pipeline — AI success starts with data quality, not model selection
2. Evaluate infrastructure alternatives — The Cerebras deal shows GPU isn't the only path to scale
3. Measure integration, not just model performance — Does AI output reach the workflows where it creates value?
4. Build operational capabilities — Monitoring, retraining, and error handling determine production success

Share-worthy stat: ”Cerebras just landed a $10B+ deal with OpenAI—the largest non-Nvidia AI infrastructure contract in history. The GPU monopoly narrative just got complicated.”

Go deeper: Track AI infrastructure trends in real-time →

The Track of the Day

”We're still in the early days of understanding how to deploy AI in production at scale.”
— DDN AI Infrastructure Report, January 2026

Like a producer who realizes mixing in the studio is only half the work—the other half is making it sound good on every sound system—the AI industry is learning that building models is easier than deploying them. The 65% aren't failing at AI. They're failing at production.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: January 15, 2026 | Curated by Yves Mulkers @ Ins7ghts