Nvidia Pulls the Plug on H200 Chip Sales to China
1. Key Themes
Theme: Decision models are a new, fast-growing AI category, and data quality is the moat
Rapid adoption of a month-old model type
- Decision models are LLMs tuned to choose between preset options cheaply and quickly, and adoption has been immediate.
- Quote: "In October, more than one in four requests on Vercel's AI model router went to decision models, not LLMs — an enormous share of work for a kind of computerized intelligence that was invented only a month earlier."
- Quote: "Decision models are LLMs that don't produce output tokens; instead, they're tuned to use their general intelligence to quickly and cheaply decide between preset options, while registering their level of confidence in the choice."
Low barriers to entry commoditize the architecture, so data differentiates
- Hyperscalers cloned the model type almost instantly, pushing differentiation toward calibration data and customization.
- Quote: "However, it's easy to make a decision model, and dozens were released in the wake of Jev. One Amazon engineer made one in a weekend, and soon every hyperscaler had its own option."
- Quote: "Data appears to be the key to a successful model in this category — TypeSafe CEO Diogo Almeida told TechCrunch that the hardest part of building Jev was finely calibrating its judgement using statistically-adjusted synthetic data."
- Quote: "Sqwish, too, has been working with synthetic data, but its unique offering is built-in post-training tools that use a customer's own data to improve the model's judgement."
Theme: Mega-rounds are concentrating in AI-for-science, robotics, and neurotech, at very high valuations
Frontier "AI plus physical world" companies are raising at billion-dollar-plus scale
- Quote: "Isomorphic Labs, a five-year-old London startup that builds AI models that simulate chemistry to design drug candidates, raised a $5 billion Series C."
- Quote: "Genesis AI, a two-year-old Paris startup that develops general-purpose robots that perform multi-step tasks in manufacturing, logistics, and laboratories, raised a $500 million Series A round at a $3.5 billion post-money valuation."
- Quote: "Science, a five-year-old startup based in Alameda, CA, that develops wearable and implantable brain-computer interfaces that restore vision for people with degenerative eye diseases, raised a $400 million Series D round at a $3.5 billion post-money valuation."
Late-stage fintech keeps stepping up valuations quickly
- Quote: "Ramp... is reportedly in early talks to raise $1 billion+ at a valuation of $70 billion+, up from $60 billion last month."
Theme: The Anthropic IPO is a landmark liquidity event, with tight hyperscaler entanglement
Demand is pushing pricing above range
- Quote: "Anthropic plans to raise the price of its IPO above the initial $70-to-$78 range after strong investor demand, with advisers discussing a price above $80 that could value the AI company at $1.5 trillion to $1.6 trillion ahead of its Nasdaq debut Thursday."
Amazon and Google hold a large stake in a company that is also their customer and competitor
- Quote: "Amazon and Google will own more than 28% of Anthropic after its IPO and concurrent private placement, underscoring the tight financial ties between the AI company and the cloud providers that are also among its biggest investors, customers, distribution partners, and competitors."
Theme: Nvidia faces a leadership and geopolitical reset
Export-control posture is tightening post-Huang
- Quote: "Nvidia is ending sales of its H200 chips in China, abandoning a controversial arrangement brokered by Jensen Huang and President Trump as the company tightens its approach to export controls following Huang's ouster."
The board is resisting a founder comeback despite external pressure
- Quote: "Nvidia said today that it does not plan to reinstate Jensen Huang as CEO despite his offer to return, with chairman Harvey Jones saying the board unanimously accepted Huang's resignation and is considering both internal and external candidates for a permanent successor."
Regulatory overhang
- Quote: "The SEC is reportedly investigating Nvidia, although the precise focus of the probe has not yet been disclosed. Possible areas of scrutiny include company disclosures surrounding Jensen Huang's relationship with Alice Huang, Nvidia's business in Asia, and its chip sales to China."
Theme: Vertical AI agents continue to attract steady seed-to-Series C funding
Agents are being applied to unglamorous, workflow-heavy industries
- Quote: "Numa... uses AI agents to handle customer conversations and automate sales and service work for car dealerships, raised a $45 million Series C round."
- Quote: "Arva AI... develops AI agents that automate financial-crime compliance and lending decisions for banks and fintechs, raised a $20 million Series A round."
- Quote: "Bevado... uses AI agents to automate leasing follow-ups, maintenance coordination, scheduling, and resident communications for property managers, raised a $10 million seed round."
2. Contrarian Perspectives
The decision-model category may be a commodity, with the moat in calibration data rather than architecture
- The article notes that the first-mover's model family still dominates usage, yet the architecture is trivially replicable. This implies model builders are less defensible than data and customization layers.
- Quote: "TypeSafe became a $7.5 billion company after releasing its Jev family of models, which continues to dominate usage (at least where we can see metrics from services like OpenRouter, which handled more than 1.5 billion queries to TypeSafe's models last week)."
- Quote: "However, it's easy to make a decision model, and dozens were released in the wake of Jev. One Amazon engineer made one in a weekend, and soon every hyperscaler had its own option."
The Nvidia board is declining to bring back its founder despite broad pressure, signaling a governance-first posture
- Consensus pressure from employees, investors, tech leaders, and the President favors Huang's return, but the board is moving on with a formal search.
- Quote: "Jensen Huang said he would be 'honored' to return as Nvidia CEO if the board asks him as employees, investors, tech leaders, and President Trump continue pressing for his comeback."
- Quote: "the board unanimously accepted Huang's resignation and is considering both internal and external candidates for a permanent successor."
Frontier-AI "safety" is brittle: refusals can be bypassed by framing a scenario as a game
- This cuts against the narrative that leading labs have solved misuse guardrails, which matters for investors and operators betting on defense and government use cases.
- Quote: "Leading AI models from Anthropic, OpenAI, Google, DeepSeek, and others can be coaxed into identifying civilians, surrendering soldiers, and even schools as potential military targets, with researchers often bypassing refusals simply by telling the models the battlefield scenario was a game."
3. Companies Identified
Nvidia
- Description: AI chip maker in a leadership transition.
- Why mentioned: Ending H200 sales to China, under SEC scrutiny, and searching for a permanent CEO after Huang's ouster.
- Quote: "Nvidia is ending sales of its H200 chips in China, abandoning a controversial arrangement brokered by Jensen Huang and President Trump."
Anthropic
- Description: AI lab heading to a Nasdaq IPO.
- Why mentioned: Pricing above its initial range at a potential $1.5–1.6 trillion valuation; Amazon and Google will hold a large stake.
- Quote: "advisers discussing a price above $80 that could value the AI company at $1.5 trillion to $1.6 trillion ahead of its Nasdaq debut Thursday."
TypeSafe
- Description: Developer of the Jev family of decision models.
- Why mentioned: Dominant in decision-model usage; became a $7.5B company.
- Quote: "TypeSafe became a $7.5 billion company after releasing its Jev family of models, which continues to dominate usage."
Sqwish
- Description: UK-based AI startup behind a decision model called kettei fat.
- Why mentioned: Raised a $3M pre-seed with a differentiated post-training offering using customer data.
- Quote: "Sqwish, too, has been working with synthetic data, but its unique offering is built-in post-training tools that use a customer's own data to improve the model's judgement."
Vercel
- Description: Developer platform with an AI model router.
- Why mentioned: Provided usage data showing the share of decision-model requests.
- Quote: "In October, more than one in four requests on Vercel's AI model router went to decision models, not LLMs."
OpenRouter
- Description: Model routing/metrics service.
- Why mentioned: Provides visibility into decision-model usage.
- Quote: "OpenRouter, which handled more than 1.5 billion queries to TypeSafe's models last week."
Microsoft
- Description: Hyperscaler and cloud provider (Azure).
- Why mentioned: Rolling out its own decision model.
- Quote: "Microsoft CEO Satya Nadella said that his company is rolling out its decision model, Microsoft-Decision-1.1, for wide use on Azure immediately."
Isomorphic Labs
- Description: London AI drug-design startup.
- Why mentioned: $5B Series C co-led by Greenoaks and Sequoia.
- Quote: "raised a $5 billion Series C. The deal was co-led by Greenoaks and Sequoia Capital."
Genesis AI
- Description: Paris-based general-purpose robotics startup.
- Why mentioned: $500M Series A at a $3.5B post-money valuation.
- Quote: "raised a $500 million Series A round at a $3.5 billion post-money valuation."
Ramp
- Description: Corporate card and financial software company.
- Why mentioned: Reportedly raising $1B+ at a $70B+ valuation.
- Quote: "reportedly in early talks to raise $1 billion+ at a valuation of $70 billion+, up from $60 billion last month."
Science
- Description: Brain-computer interface company restoring vision.
- Why mentioned: $400M Series D at $3.5B post-money.
- Quote: "has raised a total of $900+ million."
Model ML
- Description: London startup deploying AI agents for bankers, consultants, and investors.
- Why mentioned: $250M Series B at $1.5B post-money.
- Quote: "raised a $250 million Series B round at a $1.5 billion post-money valuation."
Galvanize Therapeutics
- Description: Pulsed electric field medical device company.
- Why mentioned: $130M Series D.
- Quote: "raised a $130 million Series D round led by HealthQuest Capital."
Tiberius Aerospace
- Description: Defense tech builder of precision-guided artillery ammunition.
- Why mentioned: $85M Series A at a $1B+ valuation.
- Quote: "raised an $85 million Series A co-led by Lakestar and Rochefort Asset Management at a $1+ billion valuation."
Compute Desk
- Description: Financial infrastructure for pricing, trading, hedging, and clearing GPU compute.
- Why mentioned: $45M seed led by Lightspeed, an unusual market-structure bet on compute.
- Quote: "builds financial infrastructure for pricing, trading, hedging, and clearing GPU compute, raised a $45 million seed round led by Lightspeed Venture Partners."
LanceDB
- Description: Multimodal data infrastructure for AI teams.
- Why mentioned: $57.5M Series B.
- Quote: "helps AI teams search, curate, and train on large multimodal datasets without moving them between systems."
Numa
- Description: AI agents for car dealerships.
- Why mentioned: $45M Series C.
- Quote: "uses AI agents to handle customer conversations and automate sales and service work for car dealerships."
Remanence
- Description: Paris startup customizing open-weight models for multistep business tasks.
- Why mentioned: $30M seed.
- Quote: "customizes open-weight AI models to carry out specialized, multistep business tasks."
Codex
- Description: Stablecoin-based cross-border payments startup.
- Why mentioned: $30M Series A at a $130M post-money valuation.
- Quote: "helps businesses make cross-border payments and currency exchanges using stablecoins."
Daisytuner
- Description: Compiler startup optimizing AI models across chips.
- Why mentioned: $6.6M seed.
- Quote: "develops a compiler that optimizes AI models to run across GPUs and other chips."
Valkryn
- Description: DC startup building test environments for autonomous military systems.
- Why mentioned: $15M seed led by a16z.
- Quote: "builds realistic test environments that measure and score how military drones, robots, and autonomous weapons perform."
Nomadic
- Description: Munich drone startup that lands on live power lines to recharge and inspect.
- Why mentioned: $6M seed.
- Quote: "builds drones that land on live high-voltage power lines, recharge from them, and inspect grid conditions."
Waymo
- Description: Robotaxi operator.
- Why mentioned: Barred from California freeways, suspending freeway rides in all 15 U.S. cities.
- Quote: "prompting the company to suspend freeway rides across all 15 U.S. cities where it operates."
Bedrock Energy
- Description: AI-guided geothermal drilling for commercial buildings.
- Why mentioned: $24M round.
- Quote: "drills boreholes for geothermal heating and cooling systems at commercial buildings using AI-guided rigs."
4. People Identified
Jensen Huang
- Description: Nvidia's founder, ousted as CEO.
- Why mentioned: Offered to return; the board declined; his China chip arrangement was reversed.
- Quote: "Jensen Huang said he would be 'honored' to return as Nvidia CEO if the board asks him."
Colette Kress
- Description: Nvidia's interim CEO.
- Why mentioned: Charting her own course; reportedly clashed with Madison Huang.
- Quote: "despite clashing with interim CEO Colette Kress and other executives."
Harvey Jones
- Description: Nvidia chairman.
- Why mentioned: Stated the board unanimously accepted Huang's resignation.
- Quote: "chairman Harvey Jones saying the board unanimously accepted Huang's resignation."
Madison Huang
- Description: Jensen Huang's daughter and senior Nvidia employee.
- Why mentioned: Remains in a senior role amid loyalty concerns.
- Quote: "remains in a senior role at Nvidia after serving as her father's de facto chief of staff."
Diogo Almeida
- Description: TypeSafe CEO.
- Why mentioned: Explained that calibrating judgement with synthetic data was the hardest part of building Jev.
- Quote: "the hardest part of building Jev was finely calibrating its judgement using statistically-adjusted synthetic data."
Satya Nadella
- Description: Microsoft CEO.
- Why mentioned: Announced Microsoft's decision model rollout on Azure.
- Quote: "Microsoft CEO Satya Nadella said that his company is rolling out its decision model, Microsoft-Decision-1.1."
Ushnish Sengupta and Federica Freddi
- Description: Co-founders of Sqwish, met as Cambridge grad students.
- Why mentioned: Repurposed small-LLM architectures into a decision model.
- Quote: "had been developing small LLMs when Jev arrived on the scene, and they realized their architectures could be repurposed as a powerful decision model."
Jeff Bezos
- Description: Amazon founder.
- Why mentioned: Says he was heavily involved in casting the next James Bond.
- Quote: "says he was 'heavily involved' in choosing Robert Pattinson as the next James Bond."
5. Operating Insights
Build differentiation on data calibration and customer-specific post-training, not model architecture
- When a model type is trivially replicable, the defensible layer is calibration quality and customer-data feedback loops.
- Quote: "Sqwish, too, has been working with synthetic data, but its unique offering is built-in post-training tools that use a customer's own data to improve the model's judgement."
Route cheap, high-frequency decisions to small decision models instead of frontier LLMs
- Use decision models where speed and cost matter more than deep reasoning, such as production routing and classification.
- Quote: "They're useful in production environments, where fast decisions can be more important than the reasoning provided by a slower and more expensive frontier model."
Treat regulatory and operating-boundary risk as a product constraint, as Waymo's freeway ban shows
- A few edge-case failures (entering construction zones) triggered a regulator ban and a nationwide suspension of freeway rides, so edge-case safety gating can be a business-critical dependency.
- Quote: "California regulators have barred Waymo robotaxis from operating on the state's freeways after several vehicles entered active construction zones, prompting the company to suspend freeway rides across all 15 U.S. cities where it operates."
6. Overlooked Insights
AI is cutting VFX rotoscoping work by up to 75%, while threatening entry-level roles
- A concrete productivity datapoint on AI in creative production, paired with a labor-pipeline risk.
- Quote: "with new tools potentially cutting the process by as much as 75% even as artists warn they still require human cleanup and could threaten entry-level jobs."
Montage Ventures is raising a modest $75M fourth fund
- A small emerging-manager data point amid mega-round headlines, showing early-stage funds in fintech, commerce, and healthcare are still forming.
- Quote: "Montage Ventures, a 13-year-old VC firm based in Menlo Park, CA, that backs early-stage companies across fintech, commerce, and healthcare, is raising a $75 million fourth fund, according to an SEC filing."