OpenAI vs Nvidia vs Anthropic: Three AI Compute Strategies Compared
OpenAI, Nvidia, and Anthropic are pursuing three different compute strategies. Here's how each camp is positioning itself, and what it means for buyers.
What are the three AI compute strategies OpenAI, Nvidia, and Anthropic are each betting on?
OpenAI is trying to own more of the stack it runs on, building its own inference chip while still buying enormous amounts of Nvidia hardware. Nvidia is trying to remain the default supplier to everyone, regardless of who wins the model war. Anthropic is spreading its compute across multiple chipmakers and clouds so no single supplier can hold it hostage. None of this is a formal alliance. It’s three different answers to the same problem: who controls the compute that AI runs on, and what happens to you if that control shifts.
TL;DR
- OpenAI revealed Habanero, its first custom inference chip, and says it beats Nvidia’s GB200 and GB300 systems on latency and throughput per kilowatt across several open-weight model tests, while remaining a massive Nvidia customer with roughly 12 gigawatts of Nvidia systems planned or committed through 2030.
- OpenAI used its own coding models to help design and program the chip, with AI-generated code running noticeably faster than human-written code on parts of the project, and getting three new model families running on the chip in about two months.
- Nvidia’s response, delivered by Jensen Huang on its earnings call, is that it sells systems rather than a single chip, covering training, fine-tuning, networking, and inference across every major cloud, so it doesn’t need to win every workload to stay central to the market.
- Anthropic has deliberately diversified its compute, using Amazon’s Trainium chips, a multi-gigawatt Google TPU agreement built with Broadcom, Nvidia capacity via Microsoft, and now a large SpaceX AI data center, giving it the ability to shift workloads if any one supplier falls short.
- The OpenAI-Cursor split shows the real risk of depending on one vendor for both intelligence and memory: after SpaceX bought Cursor, OpenAI said it would stop supplying future models to the tool, echoing Anthropic’s earlier withdrawal from Windsurf.
- For individual users and teams, the practical lesson is to keep memory, files, and instructions outside of any single AI provider so a corporate deal or falling-out between companies doesn’t strand your work.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Why did OpenAI build its own chip if it still needs Nvidia?
OpenAI’s chip, code-named Habanero, is an inference chip. It’s built to run models cheaply and efficiently at huge scale, not to train them. OpenAI has said it beats Nvidia’s GB200 and GB300 systems, and Nvidia’s newer Vera Rubin generation, on latency and throughput per kilowatt in specific open-weight model tests. The chip reportedly went from a blank design to full tape-out in about nine months, a fast turnaround for custom silicon.
That speed came partly from using OpenAI’s own coding models during the design process. For selected parts of the work, AI-generated code ran 1.5 to 1.8 times faster than code written by human engineers. This isn’t unique to OpenAI: DeepSeek has done similar AI-assisted kernel optimization, sometimes without a full explanation of why the resulting code performs better.
None of this replaces Nvidia. Training frontier models, handling unusual or unplanned workloads, and running the bulk of OpenAI’s infrastructure still depends on Nvidia systems. Nvidia itself has said OpenAI has roughly 12 gigawatts of Nvidia systems planned or committed through 2030. Habanero is about controlling the economics of one specific, high-volume workload: the repeated inference calls that power ChatGPT and Codex billions of times over. Owning even a slice of that changes OpenAI’s unit economics without requiring it to abandon its biggest supplier.
How is OpenAI trying to own the whole AI stack, not just the chip?
The chip is one piece of a larger loop OpenAI is building. ChatGPT usage data shows OpenAI which workloads are most expensive, which informs how models get trained and optimized. Codex and other coding models improve the software that runs on chips like Habanero. That combination feeds back into how the next chip generation gets designed. OpenAI doesn’t need to run every workload on its own silicon for this to work. It just needs to own enough of the expensive, repeated inference work to shift its cost structure.
This same instinct showed up in OpenAI’s decision around Cursor. After SpaceX acquired the coding tool, Cursor began offering Grok and Composer as options alongside other models. OpenAI responded by saying it would stop providing future models to Cursor, citing trust and contractual concerns created by the new ownership. Anthropic made a similar move earlier when it withdrew support from Windsurf. The pattern is consistent: companies pull models from surfaces they don’t control once a rival owns the distribution layer, especially where there’s risk that usage data or model behavior could be extracted by a competitor.
Why does Nvidia say it doesn’t need to win every deal?
Jensen Huang’s response to custom chips like Habanero, laid out on Nvidia’s earnings call, wasn’t a claim that they’ll fail. It was an argument that Nvidia’s business isn’t built on any single chip. Nvidia sells the systems used to train models, to fine-tune them, to network thousands of GPUs together inside data centers, and to run varied, unpredictable workloads. It sells that stack through every major cloud provider, which makes it structurally difficult to route around entirely.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
This is why Nvidia can afford to invest in companies that are simultaneously trying to reduce their own Nvidia spend. Huang’s bet is that the overall AI compute market is growing fast enough that custom chips can carve out specific niches (like OpenAI’s inference workload) while Nvidia still grows in absolute terms. Google can build and sell its own TPUs while also selling Nvidia systems through Google Cloud, because it doesn’t have enough TPU capacity to go it alone. Anthropic can buy Amazon and Google chips while still relying on Nvidia capacity delivered through Microsoft and SpaceX. Every path in this market runs through Nvidia somewhere, which is precisely the position Huang is defending.
What makes Anthropic’s multi-supplier approach different?
Anthropic’s strategy is the least tidy of the three, and that seems to be intentional. It uses Amazon’s Trainium chips, has a multi-gigawatt agreement for Google TPUs built with Broadcom, gets Nvidia capacity through Microsoft, and now uses a large SpaceX AI data center (Colossus, with more than 220,000 Nvidia GPUs) to run Claude. Notably, Claude has not been pulled from Cursor, even though Cursor is owned by SpaceX, a company that competes directly with Anthropic through its own models.
This spread gives Anthropic flexibility that neither OpenAI nor Nvidia has in the same way. Different workloads can run on different chips. Suppliers compete on price and availability to keep Anthropic’s business. If one supplier can’t deliver enough capacity, work can shift elsewhere. The tradeoff is that Anthropic gives up the tight, single-loop coordination OpenAI is building between its models, its chips, and its coding tools. What it gets in return is reduced dependency on any one company controlling its compute future, which matters most when compute itself is the binding constraint on growth.
Is any one of these three compute strategies clearly the “right” one?
Each strategy solves a different problem, and none of them are mutually exclusive in practice. OpenAI’s vertical integration lowers costs on high-volume, predictable workloads while still leaning on Nvidia for everything else. Nvidia’s sell-to-everyone approach hedges against picking a losing model or chip architecture, since it profits regardless of who wins as long as the overall market keeps growing. Anthropic’s multi-supplier approach trades some efficiency for negotiating power and resilience against any single point of failure.
For companies with OpenAI’s scale and cash, vertical integration makes sense. For a chip and systems maker, staying supplier-agnostic is the safer long-term bet. For a model company without a dominant cloud or hardware arm, diversification is close to the only rational move. The more interesting question for buyers and builders isn’t which strategy is best in the abstract, but which one leaves you least exposed if the company you depend on changes hands, changes its mind, or gets into a fight with a partner.
How should individuals and teams think about compute lock-in?
The OpenAI-Cursor split is the clearest recent example of the real risk: a year of chat history, saved memories, project context, and custom instructions can sit inside a single tool, and that tool’s access to a given model can vanish overnight if ownership changes or two companies fall out. Switching tools doesn’t help much if your accumulated context doesn’t travel with you.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
The practical response is to separate memory and files from any single model provider. Keep documents in normal files you control, keep code in repositories you own, and keep instructions and workflows portable rather than locked inside one app’s proprietary memory system. Tools that route prompts across multiple models, such as OpenRouter, can help avoid dependence on one provider for raw model access, but they’re not a substitute for owning your own memory and file layer. Budget-wise, the sensible approach scales with spend: a single main provider plus a free backup account at lower budgets, direct access to two or three providers plus a coding tool in the middle range, and multiple frontier-tier subscriptions used hard enough to justify their cost at the high end.
Frequently Asked Questions
What is OpenAI’s Habanero chip used for?
Habanero is OpenAI’s first custom chip, designed for running (not training) AI models. OpenAI says it outperforms Nvidia’s GB200 and GB300 systems on latency and throughput per kilowatt in specific tests, and it’s aimed at lowering the cost of high-volume inference work like ChatGPT and Codex requests.
Why did OpenAI stop supporting Cursor?
After SpaceX acquired Cursor, OpenAI said it would stop providing future models to the tool, citing trust and contractual issues created by the new ownership. Cursor’s owner, SpaceX AI, competes with OpenAI, and OpenAI was concerned about model access inside a rival-owned surface.
Does Anthropic use Nvidia chips too?
Yes. Anthropic uses Amazon’s Trainium chips and Google TPUs, but it also relies on Nvidia capacity delivered through Microsoft and through a large SpaceX AI data center reported to include more than 220,000 Nvidia GPUs. Its strategy is diversification, not avoidance of Nvidia.
Is Nvidia worried about custom chips like Habanero?
Not based on Jensen Huang’s public comments. Nvidia’s position is that it sells full systems, training infrastructure, networking, and cloud access across every major provider, so custom inference chips taking one workload doesn’t threaten its broader business as the overall AI compute market keeps expanding.
How can I avoid being locked into one AI provider?
Keep your documents, code, and instructions in formats and repositories you control rather than inside one app’s proprietary memory system. Use a primary AI provider for daily work, but maintain a working familiarity with at least one alternative, so a vendor dispute or acquisition doesn’t strand your files or workflow.
