The data bottleneck in surgical robotics

Building a robot that can safely navigate a human body is fundamentally different from training a large language model. A language model learns from text. A surgical robot has to learn from physical experience — the give of tissue, the resistance of a calcified artery, the precise amount of force needed to manipulate a catheter without causing damage.

That kind of embodied learning normally requires thousands of hours of real clinical procedures. But in healthcare, physical bodies are scarce, tightly regulated, and slow to generate the edge cases a robot actually needs to see. A kidney stone lodged at an unusual angle or a guidewire catching on a vessel wall might only appear in a fraction of cases. Waiting for them to happen in an operating theatre is not a viable training strategy.

Nvidia thinks it has a solution. The company’s new Medical Physics Simulation framework, announced as an open-source addition to its Nvidia Isaac for Healthcare platform, treats surgical robots as physical AI systems. The idea: generate the embodied experience computationally, at scale, before a scalpel ever touches a patient.

How the framework works: classical physics meets generative AI

The framework combines two modelling approaches. Classical physics simulation handles the mechanical rules that are already well understood — how a catheter bends, how much resistance a vessel wall applies, how contact forces shift as an instrument moves through tissue.

Generative AI handles what’s harder to hand-code: visual scene dynamics learned from procedural data. That part comes through a component Nvidia calls Cosmos-H Dreams.

Together, they form what the company calls a physical AI proposition. Classical simulation gives a robot policy the physics it must obey. Generative simulation gives it the range of visual and anatomical variation it needs to generalise. Run at scale on GPUs using Nvidia’s Warp and Newton libraries, the framework can execute thousands of parallel training environments instead of one scene at a time.

Nvidia claims a benchmark running 8,192 parallel environments cut training time from over five hours to under two minutes. But that measures throughput, not clinical reliability. It says nothing about how a policy trained this way performs against incomplete imaging, delayed sensor readings, or anatomy that falls outside anything the simulation modelled.

A language model that underperforms on an edge case produces a bad answer. A physical AI system that underperforms on an edge case is operating inside a patient. The parallel-simulation approach is a real advance in how fast developers can explore failure modes. Whether those simulated failure modes match what actually goes wrong in a surgical suite is a separate question — one none of the early adopters has published answers to yet.

Who is testing the approach — and how far along they really are

The organisations Nvidia names as early adopters are applying the physical AI approach at different depths. The list is worth reading with that in mind rather than treating it as a uniform roster of deployments.

CMR Surgical and Cambridge Consultants, the Capgemini-owned engineering firm, have gone furthest on the data side. CMR has contributed close to 500 hours of anonymised clinical data from its Versius Surgical Robotic System to the Open-H Embodiment dataset. The data spans cholecystectomy, prostatectomy, hernia repair and hysterectomy procedures. The pair are using Cosmos-H Dreams to model soft-tissue interaction physics and produce patient-specific simulations.

“Open-source models allow us to build on shared knowledge, accelerating responsible innovation and, ultimately, gives us the potential to deliver more consistent care and better outcomes for patients worldwide,” said Chris Fryer, CTO at CMR Surgical.

Johnson & Johnson MedTech is using the framework alongside a Cosmos-based foundation model to build a digital twin of its endoluminal MONARCH platform, focused on kidney-stone scenarios in urology.

XCath is applying it to endovascular autonomy policy training — teaching a system the physical behaviour of navigating blood vessels without a human hand on the controls. Inner Logic is generating synthetic data to validate device mechanics and says it intends to produce in silico evidence to support regulatory submissions. No submission built on that evidence has been confirmed publicly.

Medtronic Structural Heart sits earliest in the group, exploring simulated X-ray sensing for catheter navigation research.

Each of these is a training exercise or dataset contribution. None is a deployed system operating on a patient with policies learned this way. Nvidia doesn’t claim otherwise.

Why open source matters for physical AI in healthcare

Healthcare robotics carries a governance requirement most physical AI applications — including industrial and warehouse robots — don’t face to the same degree. Regulators and clinical review boards need to see how a system arrived at its behaviour, not just confirm the behaviour looked acceptable in testing.

An open-source framework lets developers inspect the physics assumptions inside the simulation, reproduce results across different anatomies, and build an evidence trail suited to a submission before the FDA or an equivalent body.

That’s a stronger argument for openness in physical AI than it is in most software categories, where a closed vendor pipeline hides the assumptions a team would otherwise need to defend to a regulator. It doesn’t settle the validation question on its own. Open code lets outside reviewers check the model’s logic, but it doesn’t confirm the model’s physical behaviour matches what happens in a body. That confirmation still has to come from testing that none of these companies has published yet.

The road ahead: simulation as infrastructure, not replacement

Nvidia has built infrastructure that could shorten the pre-hardware phase of physical AI development for surgical and diagnostic robots. Running training at this scale in parallel is a departure from rebuilding a custom simulation scene for every workflow. For teams that previously spent months gathering enough clinical data to train a single policy, the ability to generate thousands of parallel training environments on demand is a real shift.

But simulation remains a tool for exploration, not a substitute for clinical validation. The companies testing Nvidia’s framework are still early in the process. None has published results showing a policy trained entirely in simulation performing reliably on real patients. The gap between simulated failure modes and real surgical complications is the gap that will determine whether this approach delivers on its promise.

For now, Nvidia has given the healthcare robotics community a faster way to ask better questions. The answers will take longer to arrive.

For more on how simulation is reshaping medical training and device development, see AI in drug discovery and healthcare robotics trends.

Kimi K3 lands, and the policy debate reignites

On July 16, Moonshot AI dropped Kimi K3, the largest open-weight model ever released. Within days, it had reopened a policy argument in Washington that had been dormant for a year. The question for enterprises evaluating Chinese open-weight models this month isn’t about benchmarks. It’s about whether using one will still be straightforward a year from now.

The outcome will affect procurement decisions well outside the United States. The mechanisms under discussion — federal procurement rules, export blacklists, security advisories — travel through the same cloud providers that serve most of the world.

The immediate trigger: a post by Dean W. Ball, OpenAI’s head of strategic futures and until recently a senior AI adviser in the Trump White House.

Ball’s forecast: regulatory risk, not a ban

Ball’s assessment of the model was largely positive. He called it a very good model whose performance he didn’t think could be explained away by distillation. He also noted it seemed ‘very token hungry’ and wasn’t obviously cheap to run — a useful caution, given K3 launches with maximum reasoning effort as its only setting and bills output at $15 per million tokens.

Then he predicted the Trump administration would eventually decide its best strategy was to create regulatory risk around Chinese open-weight models. Not a ban, which he called one of the dumber motifs in AI policy, but soft guidance from agencies suggesting such models may contain backdoors. ‘It needn’t be that well justified,’ he wrote. Enough uncertainty, and regulated enterprises retreat on their own.

Why Chinese open-weight models are a commercial problem first

The reaction was fierce, and it came from Americans rather than Beijing. David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, said he couldn’t tell whether Ball was confessing to a regulatory capture strategy or predicting one. Either way, weaponising regulatory uncertainty as a competitive tool should be unacceptable, Sacks argued.

He added that the leading closed labs, already a duopoly in model revenue, want the government to remove their open-source competition. Yann LeCun and Martin Casado argued that open and proprietary development can coexist. Ball later clarified he had been forecasting rather than recommending, and walked back the claim that open weights necessarily slow the field down.

Underneath the personalities is an arithmetic problem. Closed labs need revenue per token to justify the capital they are raising for data centres. Cheaper open-weight models compress that revenue without reducing how much AI gets used — the point Snorkel AI co-founder Braden Hancock put to TechCrunch. The routing data already shows the shift: open-weight models handled 29% of tokens through Vercel’s production gateway in June, up from roughly a ninth in April, while accounting for under 4% of spending.

That pressure is arriving from inside the American stack. GitHub made Moonshot’s Kimi K2.7 Code generally available in the Copilot model picker on July 1, hosted on Microsoft Azure. The Information reports Microsoft is now adding K3 to Azure and evaluating whether it can run Copilot features currently handled by OpenAI and Anthropic models, with potential inference savings of up to $600 million.

Microsoft has confirmed neither the figure nor which features. It’s an evaluation, not a deployment. But it’s the largest customer of both American frontier labs, pricing the alternative.

The security argument, taken seriously

Commercial motive does not make the security concern fake. The strongest version of it deserves stating. Open weights cannot be recalled. Once a model is downloaded and running inside thousands of organisations, no vendor can patch it, revoke it, or push a fix — a materially different risk profile from a hosted API. Model behaviour is harder to audit than model code: a fine-tune can carry biases or failure modes that no licence inspection would reveal.

NIST has previously found security vulnerabilities in DeepSeek’s open models. For regulated industries, questions about training data provenance and content handling are live regardless of where a model was built.

The counterargument is about proportionality rather than dismissal. Georgetown research fellow Sam Bresnick has argued that halting Nvidia H200 sales to China would slow Beijing considerably more than banning open models Americans want to use — targeting the input rather than the output. Ball himself conceded a version of this in his second observation, attributing China’s open-weight strategy partly to a lack of domestic compute for serving customers. That would make it an unintended byproduct of US export controls in the first place.

What is actually likely to happen

Axios reported on July 20, citing people close to the administration, that Commerce last year weighed adding Chinese AI labs to the Entity List. The NSA and the Office of the National Cyber Director considered issuing an advisory on Chinese AI lab threats. The White House considered an executive order making US companies liable for breaches if they used Chinese models. Officials concerned about stifling innovation killed all of it.

With adviser Sriram Krishnan gone and security hawks louder, the effort has revived. But the described approach is procurement rules, Entity List threats and public pressure rather than prohibition. ‘What’s actually happening is slower and more durable,’ one source told Axios. Neither the White House nor Commerce responded to Axios’s requests for comment. Politico reports Commerce will not move imminently.

Impact on buyers outside the US

For buyers outside the US, the exposure is indirect but real. A rule written for American regulated industries and federal procurement does not bind a Malaysian bank or an Indonesian telco. The hyperscalers are the transmission line.

Most enterprises in this region reach Kimi K3 through Azure, AWS or Google Cloud rather than Moonshot’s own API. If Washington makes hosting Chinese open-weight models uncomfortable enough for those providers, the model quietly leaves the catalogue in Kuala Lumpur at the same time it leaves it in Virginia.

Ball anticipated this in his own post, noting that regulators would not want to push so hard that hyperscalers stop serving Chinese models altogether. That would only drive startups toward less reputable providers. The obvious hedge is to hold your own copy. Moonshot publishes K3’s weights on July 27, and from that point the model cannot be withdrawn from anyone who has downloaded it.

But as covered previously, K3 is a difficult model to self-host. Moonshot recommends serving it across 64 or more accelerators, and the weights alone come to roughly 1.4TB. For most companies, the fallback is theoretical.

That leaves a narrower question than the headlines imply. Not whether Chinese open-weight models are safe or permitted. But whether the specific model you build on will still be in your cloud provider’s catalogue in twelve months, and what it would cost you to move if it isn’t. That’s a due-diligence question, and it’s answerable today.

For more on the technical side, see our analysis of the Kimi K3 open-weight model and its memory-focused architecture. Also explore how Washington AI policy is shaping global tech procurement.

The Two-Speed AI Economy Nobody’s Talking About

Here’s a riddle for the AI era: If companies are ditching expensive frontier models for cheaper open source alternatives, why is Anthropic still raking in more than half of all AI spending on major platforms?

That contradiction sits at the heart of a provocative new argument from Decagon CEO Jesse Zhang. In a post titled “Everyone is wrong about open source AI in the enterprise,” Zhang proposes that frontier labs and open source models aren’t really competing. They’re playing different roles in a single lifecycle.

Expensive frontier models handle the messy, high-risk early stages of a new use case. Once the process is proven and predictable, companies hand it off to leaner, cheaper open source models. The result? Frontier spending barely dips, because new discovery projects keep popping up to replace the ones that mature.

“The frontier labs will keep owning discovery,” Zhang writes. “Open source will increasingly own production.”

What the Data Actually Shows

Zhang doesn’t offer hard numbers, but the data is easy to find — and it largely backs up his thesis.

Take Vercel‘s AI gateway dashboard. Over the past week, DeepSeek has surged to the lead in token volume, processing just over a third of all tokens flowing through Vercel’s infrastructure. Z.ai, the lab behind the popular GLM-5.2 model, jumped to fourth place in the same period.

But scroll down to spend, and the picture flips. Anthropic still accounts for more than half of all AI spending on the platform. That share has slipped slightly — partly because Anthropic raised prices — but hasn’t collapsed.

OpenRouter tells a similar story across a broader, slightly less enterprise-focused slice of the market. DeepSeek V4 Flash dominates by raw usage, processing 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion. But the price gap is enormous: Opus costs roughly 23 times more per token ($1.37 per million tokens versus DeepSeek’s 6 cents). That means Opus likely still captures the majority of actual dollars spent.

And that’s before factoring in Nvidia’s Nemotron, which is poised to leapfrog competitors thanks to Nvidia’s deep enterprise relationships and the model’s extreme adaptability.

Why Frontier Labs Aren’t Panicking

The numbers don’t fully prove Zhang’s lifecycle theory, but they do explain why Anthropic isn’t sweating the open source surge — at least not yet.

One reason: the total pool of AI-addressable problems is expanding so rapidly that frontier labs can maintain their position simply by dominating new, unproven use cases. Every time a mature workflow migrates to open source, a fresh batch of harder problems appears to take its place.

Another explanation: some use cases are genuinely too difficult for lighter models. Even as clients experiment with cheaper alternatives, they keep a foot in the frontier door for the toughest tasks. That creates a sticky, high-margin revenue base that open source models can’t easily erode.

What This Means for Enterprise AI Buyers

For companies building on AI, the implication is clear: don’t treat frontier and open source models as an either/or choice. Use frontier models to explore and validate. Once the process is stable, switch to open source for production. It’s a hybrid strategy, not a migration.

This two-tiered economy could become a stable feature of the AI market. Frontier labs keep the premium pricing they need to fund R&D. Open source models get the volume that drives ecosystem growth. And enterprises get a cost-effective path from experimentation to deployment.

As recently as last September, many analysts — including this one — predicted that foundation labs would end up as commodity providers, selling “coffee beans to Starbucks” while the application layer captured the value. Some of that prediction came true: vertical AI startups did switch to lighter models, and the economics of “GPT wrapper” companies have remained stable.

But we’re also seeing that frontier providers have held onto the most desirable part of the marketplace: the premium token price. And that doesn’t look likely to change anytime soon.

For a deeper look at how companies are balancing cost and capability, check out our analysis of enterprise AI adoption strategies. And for more on the specific models driving this shift, see our breakdown of how DeepSeek is reshaping the open source AI landscape.