Start worrying about your dependency on specific AI companies

Cursor’s maker Anysphere announced a $2.3b Series D in November 2025 - $29.3b post-money valuation, past a billion in annualised revenue and accelerating. Bloomberg reported it topped $2b by February 2026. By any normal measure, this was one of the great startup growth stories of the decade. However, the cost structure underneath it was not such good news.

A journalist called Ed Zitron reported that Cursor’s cloud spend more than doubled in a single month: $6.2m to $12.6m by June 2025, around the time Anthropic introduced Priority Service Tiers. This points to something that goes well beyond one company’s P&L. The irony here is that Anthropic launched Claude Code - a direct competitor to Cursor’s core product - powered by the same models Cursor was paying to access.

Your biggest vendor, funded in part by your revenue, building the thing that competes with you. The dependency risk wasn’t theoretical anymore.

Cursor’s response was very smart, and very impressive. We should all be learning from them. Cursor developed its own in-house models. By November 2025, Anysphere claimed its models generated more code than “almost any other LLMs.” In March 2026, it launched Composer 2 at one-tenth of Anthropic’s Opus 4.6 list pricing, and began reducing its dependence on a single vendor.

Could your company have done that? Anysphere had a billion-dollar revenue run rate and the engineering team to pull it off. Most companies building on AI vendor APIs have neither. That asymmetry is what makes this a structural problem rather than a Cursor problem.

We’re all buying AI at prices shaped by subsidy. That won’t last.

I want to walk through the economics here carefully, because the numbers are large enough that it’s worth being precise about where they come from.

The Information reported that OpenAI expects $14b in operating losses for 2026 alone, with cumulative losses of roughly $44b before profitability arrives around 2029. Separately Reuters reported that OpenAI plans to spend roughly $200b on compute infrastructure through the end of the decade. That’s capex, not opex, but together they paint a picture of a company whose cost base is still accelerating. Reuters also reported that OpenAI’s adjusted gross margin fell from 40% in 2024 to 33% in 2025 as inference costs quadrupled.

Anthropic’s trajectory looks similar. The Information reported that gross margins improved in 2025 but remained below the company’s own targets. Anthropic has publicly discussed multi-gigawatt infrastructure commitments with Google and Broadcom scaling from 2027. Zitron reported that through September 2025, Anthropic’s AWS spend alone exceeded its estimated total revenue. If that’s accurate, and Anthropic also runs significant workloads on Google Cloud, the true cost base may be substantially larger than the reported figures suggest.

Whatever the numbers the story is pretty clear: neither company is profitable, both are raising capital at extraordinary scale, and current AI pricing is still being shaped by investment subsidy and unsettled unit economics rather than sustainable margins. That gap between price and cost appears material. It is unlikely to persist unchanged.

Per-token costs fall. Per-task costs don’t.

If you don’t agree with this, it’s really worth thinking hard about.

The cost DEflation is real and dramatic. Stanford’s AI Index 2025 report found that querying a model at GPT-3.5 performance levels dropped from $20 per million tokens in November 2022 to $0.07 by October 2024. The temptation is to extrapolate and conclude that AI costs are simply going to zero.

The problem is that per-token cost is the wrong unit of analysis for companies or professionals. What matters is per-task cost - and that’s moving in the opposite direction. Gartner’s March 2026 analysis put numbers on this: their analyst Will Sommer made the point very well - it’s not just that an agentic assistant does more queries, it’s that every single query consumes five to thirty times more tokens. As companies move from pilot chatbots to production agentic workflows, token consumption is rising faster than token costs fall.

So the per-token price is falling, the per-task cost is flat or rising, and the total bill is going up. I keep coming back to this because it’s the dynamic that most conversations gloss over.

The cost declines aren’t evenly distributed, either. The dramatic drops are concentrated on older, commodity-grade models. Frontier reasoning capabilities - the ones businesses are using - remain scarce and expensive. Both OpenAI and Anthropic have introduced tiered pricing structures with priority access, batch queues, and explicit multipliers. If your business runs primarily on commodity-grade AI, the deflation story works in your favour. If it depends on the frontier models for anything that genuinely matters, the economics look quite different.

The real risk isn’t price. It’s concentration.

I am NOT saying “AI prices will skyrocket”. Prices per token probably won’t, I think they’ll keep falling in nominal per-token terms. But that just hides the risk right now.

What genuinely worries me is the compounding nature of AI vendor dependency. Every month a company runs production workloads on a frontier API, the switching costs grow - and they grow in ways that are hard to see from the outside. Prompts get more model-specific, agent architectures (and even harnesses) couple to particular model behaviours and failure modes, evaluation infrastructure gets calibrated to one vendor’s outputs. No technical team I’ve spoken to so far has quantified these switching costs, but they’re real and they accumulate with every sprint.

This creates several vulnerabilities that compound, not diversify, the risk.

Capacity risk. GPU procurement cycles remain long and volatile - lead times for H100-class systems were reported at 36 to 52 weeks back in 2023, and while availability has improved for some configurations, forward orders from hyperscalers absorb a large share of premium capacity. Scarce compute tends to favour hyperscalers’ own workloads and their largest customers first. You can see the signals in the vendor documentation itself: priority tiers, batch queues, explicit “overloaded_error” status codes.

Supplier-as-competitor risk. The Cursor case is an early example of what I suspect will become a common pattern. Anthropic launched a competing product built on the same models its largest customer was paying to access. OpenAI embeds its models into Microsoft products that compete with its API customers. When the company selling you inference is also building the application layer on top of it, the incentive structure is not aligned in your favour.

Quality degradation and repricing risk. As vendors introduce tiered pricing, the cheap options quietly get slower, less reliable, or subtly less capable. Your processes don’t fail dramatically - they degrade gradually, which is in some ways worse because it’s harder to detect. And the effective cost of reliable, fast, frontier-quality inference rises even when the headline per-token price stays flat.

Bargaining power risk. When AI is embedded in your operations and you have no tested fallback, you have no leverage. Full stop. You don’t negotiate well when walking away means your critical processes stop working.

There’s one more layer here that I think deserves attention. A lot of companies believe they’ve managed their vendor risk because they use Claude for some tasks, GPT for others, maybe Gemini for a third workload. That feels diversified. But if those providers all run on the same two or three hyperscaler platforms - and a large share do - the concentration has just moved down a layer. The real dependency isn’t on the model vendor, it’s on the compute substrate underneath, and the hyperscaler that controls it has its own priorities when resources are limited.

The escape route exists — but it requires real investment

The gap between frontier and open-weight models has collapsed faster than almost anyone predicted. Epoch AI’s Capabilities Index analysis found that open-weight models now trail the closed-weight state-of-the-art by roughly three months on average. For many corporate tasks, that means the capability you’re paying premium prices for today often reaches deployable parity on infrastructure you control within a quarter. Bluntly - I’m using open models for ~60% of my token use, and Claude / Codex for the other 40% ish.

But - and this is important - not without real engineering overhead. Latency, reliability, and operational burden tend to be worse than managed APIs, at least initially. Running your own inference stack means building serving infrastructure, monitoring pipelines, evaluation frameworks, and hiring the engineering depth to keep it all running at production quality. I don’t know any non-tech companies that have this today. The talent market for this work is tight and competitive.

The alternative, though, is worse - discovering your dependency at the moment your vendor changes terms. Every month on frontier APIs without fallback paths deepens the lock-in. Prompts become more model-specific, harness dependency increases, switching costs compounds.

Cursor began reducing its exposure because it invested in alternatives early. Companies that wait until they feel the squeeze will find the exit much more expensive and much less certain.

What to do about it

None of this is an argument against using AI. It’s an argument for treating it with the same supply-chain discipline you’d apply to any other critical input to your business.

Segment by capability and criticality. Not every task needs frontier inference. Divide your AI workloads into four sections based on whether they require frontier capability and whether they’re business-critical. The goal: make sure the business-critical set is never single-sourced from scarce frontier capacity. In practice, this means intelligent routing - open-weight models handling the bulk of requests, frontier escalation for the genuinely hard stuff - which requires a task registry, routing policy, and per-workflow cost or token budgets. Most companies don’t have any of these yet.

Build graceful degradation paths. Every critical AI-dependent process should be able to step down: frontier API to smaller hosted model to self-hosted open-weight to deterministic rules or human fallback. If your only plan for an API outage is to wait for it to come back, you don’t have a plan. Each workflow needs a fallback ladder that’s documented and actually tested.

Treat AI capacity like procurement. If inference is embedded in your operations, it’s productive capacity, not a software subscription. Explore reserved compute agreements. Maintain relationships with multiple providers. Quantify your fallback throughput. Someone in your organisation needs to own the AI supply chain the way someone already owns your cloud infrastructure.

Measure dependency in operational terms. Stop asking “how much are we spending on AI?” and start asking “if our primary provider’s capacity were cut by 50% for 30 days, which processes fail, how fast, and at what cost?” If you can’t answer that, you don’t understand your exposure.

Start building internal capability now. For many enterprise tasks, the capability you’re paying premium prices for today will be deployable on your own infrastructure within months. The companies investing now in evaluation frameworks, deployment infrastructure, and the talent to run open models will have options. The rest will discover their dependency at the worst possible moment.

Where this is heading

AI supply chains are going to become a board-level function within the next couple of years - for the same reason cybersecurity did, which is that the operational exposure gets too large and too concentrated to sit as a line-item in someone’s IT budget.

The risk isn’t that AI becomes expensive. It’s that it becomes constrained, tiered, and strategically controlled - after you’ve already built your operations around it.