Here is my prediction:
Anthropic will be the first major frontier-model provider to fail.
By fail, I do not necessarily mean bankruptcy. Anthropic may fold. It may be acquired under pressure. It may lose enough market share that Claude becomes a niche product instead of a frontier competitor. The outcome matters more than the form.
I also do not think this will take a decade. By June 2028, I expect Anthropic’s decline to be measurable. Its share of the business AI coding market will fall from approximately 50% to approximately 20%.
That is a specific prediction, and it may not age well. Predictions should carry that risk.
Anthropic is not weak today. It is one of the strongest AI companies in the world. Claude is excellent. Claude Code has become the standard against which other coding agents are measured. Anthropic has enormous financial backing, deep enterprise adoption, and some of the best researchers in the industry.
According to the Ramp AI Index, Anthropic accounted for 41.02% of business AI subscriptions in May 2026, ahead of OpenAI at 39.54%. In the business AI coding category, CNBC reports that Anthropic controls approximately half the market.
I am predicting a massive decline from a position of strength. That is what makes the prediction interesting.
Compute Is Oil
Frontier AI companies look like software companies. They are not. They are energy and infrastructure companies with software interfaces.
Models require chips. Chips require data centers. Data centers require networking, cooling, land, and extraordinary amounts of electricity. The intelligence may appear in a chat window, but the product begins with industrial capacity.
Compute is the oil of the AI economy.
The model providers are oil companies. Some own the wells. Some control wells through partnerships and concessions. Some buy every barrel from someone else. That distinction will determine which companies survive.
Google owns data centers, designs TPUs, develops Gemini, distributes it through its products, and has the balance sheet to keep building. Meta owns enormous infrastructure, develops custom silicon, trains its own models, and can fund the work through one of the most profitable advertising businesses ever created. SpaceX owns Colossus, controls its hardware roadmap, develops Grok, is building its own low-level software stack, and is acquiring Cursor.
OpenAI does not own every building that runs its models, but it is an equity partner in Stargate and has operational responsibility for the project. Stargate is designed to build $500 billion in infrastructure for OpenAI. OpenAI has meaningful influence over how that capacity is financed, designed, built, and operated.
Anthropic is different.
Anthropic buys compute from AWS, Google, CoreWeave, Cerebras, and now SpaceX. AWS is Anthropic’s primary cloud and training partner. Google, Amazon, and SpaceX all have their own model ambitions.
Anthropic has diversified its suppliers. It has not gained control over the wells.
The Petro-States of AI
The infrastructure providers are not ordinary vendors. They are the petro-states of AI.
Microsoft, Google, Amazon, Oracle, and SpaceX control access to the resource every model provider needs. They decide what gets built, where it gets built, how quickly capacity expands, what the capacity costs, and which customers receive priority. Several also produce competing models.
That makes the relationship fundamentally different from buying office software or renting commodity servers. Anthropic is purchasing its most important strategic input from organizations that may benefit if Anthropic loses.
The petro-states are happy to sell oil while they have excess production. The incentives change when their own industries need every available barrel.
Google can prioritize Gemini. Microsoft can prioritize its own models or its preferred partners. Amazon can prioritize Nova. SpaceX can prioritize Grok and Composer.
Contracts provide protection. Diversification provides protection. Long-term capacity agreements provide protection. None of these provide ownership. Control granted by contract lasts until the contract changes, expires, becomes uneconomic, or collides with a larger strategic interest.
Anthropic’s model weights, research, brand, and customers remain valuable. Without enough competitively priced compute, those assets cannot produce a frontier product. An oil company without oil is a brand and a collection of engineers. An AI company without compute is much the same.
OpenAI Is Next
Anthropic is the most exposed frontier-model provider. OpenAI is next.
OpenAI’s position is stronger because it has an ownership and operating role in Stargate. The capacity is being built for OpenAI, and OpenAI helps determine how the system is designed and operated. That is meaningful control. It is not complete ownership.
Oracle, SoftBank, Vantage, SB Energy, Microsoft, CoreWeave, NVIDIA, and other partners still finance, build, own, or supply major parts of OpenAI’s infrastructure. OpenAI can direct the wells, but it does not own all the ground beneath them.
They resemble an oil company operating through concessions and joint ventures. The company may control production for decades. It may invest billions in equipment, employ the engineers, direct operations, and depend on the output. The underlying assets still sit inside relationships with governments, financiers, and infrastructure owners whose priorities can change.
OpenAI has more protection than Anthropic because it sits inside the governance and operation of its infrastructure program. Anthropic remains primarily a customer purchasing capacity across several providers.
That creates a spectrum:
- Anthropic rents the wells.
- OpenAI controls wells through partnerships.
- Google, Meta, and SpaceX own the land and every piece of the supply chain.
If control over compute determines the order of failure, Anthropic goes first, and OpenAI is next. The vertically integrated providers inherit the advantage.
Anthropic Is Financing a Competitor’s Wells
The SpaceX agreement makes this risk concrete.
SpaceX signed an agreement to provide Anthropic access to Colossus 1, a cluster containing more than 220,000 NVIDIA GPUs. The capacity can support training, fine-tuning, inference, and high-performance computing. Anthropic said the immediate capacity would improve service for Claude Pro and Claude Max customers.
According to TechCrunch’s analysis of SpaceX’s S-1 filing, Anthropic secured the entire 300-megawatt output of Colossus 1 for $1.25 billion per month through May 2029. The arrangement could pay SpaceX more than $40 billion. Either party can terminate with 90 days’ notice.
At first glance, this is simply a large customer buying capacity from a supplier with unused infrastructure. Anthropic needs compute. SpaceX has compute. SpaceX receives approximately $15 billion in annual revenue, and Anthropic expands Claude capacity.
The arrangement makes sense today. Strategy is about what happens when today’s incentives change.
Anthropic is not merely buying oil from a competitor. It is financing the competitor’s wells.
The revenue helps SpaceX improve Colossus, expand its infrastructure, recruit engineers, optimize its stack, and train the models that will compete with Claude. Anthropic’s demand supports the industrial base SpaceX can eventually turn against it.
SpaceX does not need to terminate the agreement tomorrow. It can collect the revenue while it builds the alternative. When the alternative is ready, SpaceX’s incentives change.
Then SpaceX Bought the Gas Stations
Compute alone is not enough to take Anthropic’s market. SpaceX also needs distribution.
That is what Cursor provides.
SpaceX has entered a merger agreement to acquire Cursor at an implied equity value of $60 billion. The stock transaction is expected to close in the third quarter of 2026, subject to regulatory approval.
There is some symmetry here. Cursor once accounted for roughly 40% to 50% of Anthropic’s revenue. Anthropic reportedly told Cursor that Claude Code was a research effort rather than a serious commercial push. Then Claude Code became a direct competitor and began taking Cursor’s market share.
Anthropic used Cursor as a distribution channel, then competed with it. If SpaceX uses Cursor to supplant Claude Code, Cursor will return the favor. Anthropic opened a competing gas station. Now the original gas station belongs to the company that owns the wells.
Cursor gives SpaceX an established coding product, enterprise customers, developer mindshare, usage data, and a place to distribute its models. It also gives SpaceX Composer, Cursor’s own model effort.
According to CNBC’s reporting on the transaction, Cursor’s market share fell from 41% in June 2025 to approximately 26% in May 2026. Anthropic now controls roughly half of the category.
That does not make Cursor irrelevant. It explains the acquisition.
SpaceX is buying a major distribution channel after Anthropic took the lead. It is not entering an adjacent market for diversification. It is acquiring the product layer needed to attack Claude Code. The oil company is buying gas stations.
That matters because customers may not need to make a visible switch. Cursor can make Composer or Grok the default model while preserving the interface developers already use. The model changes underneath the product.
Anthropic can lose model share without every customer holding a migration meeting. Distribution can move the market for them.
SpaceX Is Building a Better Well
SpaceX’s advantage is not limited to owning more GPUs.
Elon Musk says SpaceX has nearly finished an in-house AI training stack written primarily in C. The software maps directly to 220,000 GB300s with 800-gigabit networking, makes heavy use of pipeline parallelism, and gets as close to bare metal as possible.
Musk claims the potential speed improvement over JAX for large training runs exceeds an order of magnitude. That claim has not been independently demonstrated. It does not need to be true for the strategy to work.
In a follow-up, Musk said SpaceX will build an inference stack in C next for simultaneous high-speed reinforcement learning across a large block of GB300s.
The training stack is nearly complete. The inference stack is announced work. Based on SpaceX’s record of delivering difficult infrastructure, I expect the inference work to ship.
I do not need to assume a tenfold advantage. SpaceX only needs a meaningful one.
If tighter integration lets SpaceX extract more training work and more inference from every GPU and megawatt, it lowers the cost of producing intelligence. In oil terms, SpaceX is reducing the lifting cost of every barrel.
Anthropic pays infrastructure providers. SpaceX controls the well, the pump, the refinery, and the gas station. Its C stack is designed to pump more from each well, refine it faster, and deliver more finished product from the same infrastructure.
That vertical integration gives SpaceX room to compete on price even before it produces a model that is clearly better than Claude.
Parity Is Enough
SpaceX does not need to beat Anthropic. It only needs to make paying a premium for Anthropic difficult to justify.
I lead software development teams. If I can get Sonnet-level coding performance for 20% less, I will switch.
Today’s AI subscriptions hide the marginal price of heavy consumption. Anthropic limits five-hour sessions and weekly usage. OpenAI Codex uses five-hour windows, additional weekly limits, and credits. GitHub Copilot provides monthly AI-credit allowances.
These are not unlimited subscriptions. They are rationed access to an expensive resource.
My team already works across those boundaries. GitHub Copilot is our primary tool, but our highest-demand users also have Claude Code subscriptions to keep costs under control. We use the included consumption from two companies, and developers move between the tools without disrupting the work.
I will not need a six-month migration plan.
I will need two days.
My repositories remain where they are. My developers keep their workflows. My stories, standards, review process, and deployment pipeline do not change. The team adjusts prompts, updates configuration, runs evaluations, and learns the small behavioral differences in the new model.
There is friction, but it is minimal compared with changing a database, cloud platform, programming language, or core business system.
Model providers have weak switching moats. The customer’s code, data, business logic, and process live outside the model. Coding agents and model routers make substitution easier. Enterprise security reviews, negotiated contracts, prompt behavior, and model-specific tooling add resistance, but they do not create the kind of lock-in enterprise software companies traditionally depend on.
Compute is scarce. Customer loyalty is scarce, too. It lasts only while the supply remains reliable, capable, and affordable.
The competitive bar is not superiority. It is acceptable parity at a better price.
SpaceX can also bundle. It can price Composer or Grok aggressively inside Cursor. It can subsidize model usage to rebuild Cursor’s share. It can optimize the product and model together. It can trade margin at one layer for adoption at another.
Anthropic has to defend model margin while paying its petro-states. That is not an equal cost structure.
What “Nationalization” Looks Like in AI
The oil industry provides the failure mechanism. Venezuela nationalized its oil industry in the 1970s. Foreign operators could possess expertise, contracts, equipment, and customers. The state controlled the resource underneath them.
When the state changed the rules, contractual control was not ownership. That is where the analogy applies. In AI, no government needs to seize a data center. The infrastructure provider already owns the resource and can decide that its compute is more valuable powering its own products than those of a customer that has become a competitor.
The contract expires. The price rises. Expansion slows. The provider reserves its newest hardware for its own models and redirects capacity toward its own strategic goals. The customer receives yesterday’s capacity while the owner builds tomorrow’s product.
Nothing has been seized. The owner is exercising ownership. The former customer simply discovers that dependable access to a resource was never the same as control over it.
Anthropic would still own Claude. It could still have customers, revenue, researchers, and valuable intellectual property. It would no longer control enough cost-effective production to remain at the frontier.
A model without compute is an oil company after the well owner redirects production elsewhere.
The Strongest Case Against This Prediction
Anthropic is not dependent on one supplier. That is the strongest objection, and it is true.
Anthropic has relationships with AWS, Google, CoreWeave, Cerebras, and SpaceX. It has committed access to multiple hardware architectures. It can shift workloads. It has raised extraordinary amounts of capital. Its current growth gives infrastructure providers every reason to keep selling it capacity.
SpaceX also has no immediate reason to abandon $15 billion in annual compute revenue. Cursor’s acquisition is not yet closed. Composer and Grok have not demonstrated Sonnet-level performance at a meaningfully lower production cost. Musk’s performance claims remain projections.
Anthropic is also reportedly moving toward leasing and managing data centers directly. That would give it more operational control and reduce part of the exposure described here.
Those facts could prove me wrong. They do not remove the structural problem.
Every major Anthropic compute relationship still depends on someone else’s capital, facilities, chips, power, and strategic priorities. Diversifying among petro-states reduces the risk that one supplier cuts you off. It does not turn you into a petro-state.
Anthropic is trying to secure more wells through longer contracts and more suppliers. Its competitors are building and controlling the wells.
The difference will matter most when compute becomes even more constrained, model quality converges, and price competition intensifies. I expect all three to happen.
How to Know If I Am Wrong
Here is how to check me.
By June 2028, Anthropic’s decline will be visible through sustained loss of business AI coding market share. I expect its share to fall from approximately 50% today to approximately 20%.
Anthropic will be the first major frontier-model provider to fold, be acquired under pressure, or become a niche product rather than an independent frontier competitor.
The signals will appear before the outcome:
- Cursor will shift more usage toward Composer, Grok, or other non-Anthropic models.
- SpaceX will deliver a competitive coding model and price it below Claude.
- SpaceX’s C training and inference work will produce a meaningful cost or speed advantage.
- Anthropic’s market share will fall faster than its overall market grows.
- Anthropic will spend more aggressively to secure capacity it still does not strategically control.
- Infrastructure providers will reserve their newest or most efficient capacity for their own models and preferred partners.
If Anthropic still controls approximately half of the business AI coding market in June 2028, the prediction failed. If Anthropic builds or controls enough infrastructure to erase the cost disadvantage, the prediction failed. If Composer and Grok remain materially behind Claude, the prediction failed. If switching costs prove stronger than I expect, the prediction failed.
That is the value of writing predictions down. You do not get to remember only the parts you got right.
Anthropic has built an extraordinary refinery.
It may be the best refinery in the market today.
It still does not control the wells.
Sources and Measurement
The market-share figures in this prediction measure different but related markets. The Ramp AI Index measures paid business AI subscriptions across more than 70,000 businesses. CNBC’s approximately 50% figure refers specifically to the business AI coding category. I use the narrower coding figure as the baseline for the June 2028 prediction.
The compute agreement is documented by SpaceX, with the financial terms reported by TechCrunch from SpaceX’s S-1 filing. The Cursor transaction is documented in SpaceX’s SEC filing.
The claims about SpaceX’s C training and inference stacks come directly from Musk. The software, performance advantage, pricing response, and resulting market-share shift remain predictions, not established facts.
The subscription-usage examples come from Anthropic’s usage-limit documentation, OpenAI’s Codex pricing documentation, and GitHub’s Copilot plan documentation. Those limits establish that access is rationed. The conclusion that this weakens loyalty and increases price sensitivity is my inference.