The Foundation Model Trap: A Wake-Up Call For Biopharma

Over the last couple of weeks, the conversation around enterprise AI shifted. If you caught the recent All-In episode where Sacks, Chamath, and the crew unpacked Alex Karp's on-air blowup at the frontier labs, you heard the quiet part said out loud. The honeymoon phase with the AI labs is over.

It's been replaced by a hard realization in the C-suite: enterprises aren't just buying software anymore. They're funding their future competitors.

The original premise of the AI boom was clean. OpenAI and Anthropic would be the new utilities. They'd build the raw intelligence — the base models — and the rest of the world would build applications on top of them. Plumbing below, businesses above. Everybody wins.

It isn't working out that way. The labs have figured out that selling raw tokens is a low-margin, brutally competitive race to the bottom. To justify their valuations, they can't just be the plumbing. They have to capture the value at the application layer.

And that means they have to enter your business.

If you run R&D in biopharma, this isn't theoretical or on the horizon — it landed on June 30, and I'll get into exactly what happened. But to see why it should worry you, you have to see the pattern it belongs to. That pattern is already out in the open in design and in code, where the labs moved first. So I'll start there, and follow it home to pharma.

The Venture Trap and the Growth Imperative

To understand why a company like Anthropic is dangerous to build on, you have to look at the venture math behind it.

In May, Anthropic raised $65 billion at a $965 billion valuation — nearly tripling its worth in three months, on run-rate revenue that had just crossed $47 billion. Sit with those numbers. You do not grow into a near-trillion-dollar company charging fractions of a cent per thousand tokens. The math doesn't close.

The only thing that closes it is vertical integration. At some point, every lab looks at its highest-paying customers and asks the predatory question: Why are we letting them keep the margin?

This is why the enterprise has to move toward platforms that are strictly model-agnostic — platforms built to orchestrate your data, not to chase application-layer revenue. When your infrastructure vendor doesn't build models, their incentives line up with yours. They win when your data stack wins.

Anthropic's incentives are the opposite. They win when they own the whole stack. And if you're building on top of them, you're handing them the roadmap to replace you.

The Figma Case Study: From Partner to Predator

If you want the pattern in one clean example, look at Figma.

Early in 2026, Figma and Anthropic were close partners. They were building a feature together — "Code to Canvas" — that let Claude work natively inside Figma. Figma was a premium customer, routing enormous volumes of data and proprietary design data through Anthropic's APIs. Anthropic's chief product officer, Mike Krieger, sat on Figma's board.

Then the growth imperative kicked in.

On April 14, Krieger resigned from that board — the same day The Information reported Anthropic's next model would ship design tools. Three days later, Anthropic launched Claude Design, a standalone tool built on Claude Opus 4.7 that turns a text prompt into working prototypes and slides. Figma's stock dropped between 6 and 7.7% that day.

Board seat to competitor in seventy-two hours.

Design isn't the only front. In coding, Anthropic ships Claude Code at a price that undercuts Cursor — the editor built largely on Anthropic's own models — and subsidizes it so aggressively that a single $200 subscription can burn up to $5,000 in compute. The infrastructure provider and the application competitor are the same company, and it can afford to lose money to take the market.

This isn't a pivot. It's a playbook. They use your API spend to fund their research, watch how your users work to learn your workflow, then ship the product that cuts you out. So ask the honest question: how do you build a ten-year strategy on top of a vendor that runs this play?

The Karp Thesis: The Wealth Tax on Your Alpha

Alex Karp said the quiet part on CNBC on July 1, and he didn't soften it. Enterprises, he said, are "livid." He called closed-model token billing a "wealth tax" — one that "does not help the poor, it just punishes." His read on what customers tell him privately: "I am paying for tokens that create no value. These people are stealing the weights and alpha of my business." His verdict on the whole model: "Something has gone completely wrong."

You pay this tax twice.

The Financial Tax. You pay premium rates for high-volume reasoning tokens, and the spread is brutal. On All-In, Chamath laid out the math: for a billion input and a billion output tokens a month, GPT-5.5 Pro runs about $105,000. Claude Opus 4.8, about $30,000. DeepSeek V4 Pro, about $5,220. DeepSeek R1, about $2,740. That's a 10-to-30x spread. And his sharper point — the capability gap between the best open and closed models is closing far faster than the pricing gap. You're paying a fortune for a lead that shrinks every quarter.

The Intelligence Tax. You hand over your alpha — the domain logic, the proprietary data, the workflows that make your business yours. Even if a lab swears it isn't training on your prompts, your usage patterns reveal how valuable software gets built in your vertical. They learn your business on your dime. And "we won't keep your data" is a promise, not a property of the system. When Anthropic shipped its Fable 5 model class in June, it forced mandatory 30-day retention on every prompt and output — no opt-out, no enterprise carve-out — voiding the zero-retention contracts customers had already signed. Microsoft barred its own employees from the model within a day. Your NDA is only as durable as the vendor's next release notes.

David Sacks put the defense best on that same episode. For an enterprise, he argued, AI safety has nothing to do with the headlines. It means not losing the means of production — control over your compute, your models, your data stack, and your alpha. Hand those to a lab, and it can absorb your know-how and turn it into a product that competes with you. He said Karp was "exactly right." The moment a lab sees how much value you're generating in a niche, it has every incentive to verticalize into it.

The Switch: They Decide Which Model You Get

Here's the part that should end the debate for anyone still calling these labs neutral utilities. It's not that Anthropic might compete with you. It's that they've written down, in their own documentation, that they decide which version of the model you receive — and keep the best one for themselves.

Fable 5, released June 9, was their most capable public model. But it shipped in two forms. The full-strength version, Mythos 5, goes to Anthropic and a short list of approved organizations. The governed, monitored version, Fable 5, goes to you. Same model underneath. The vendor decides who sits on which side of the line.

Then it gets specific. When Fable detects a request touching cybersecurity, biology, chemistry, or model distillation, it hands your response to the older, weaker Opus 4.8 — and tells you it did. Argue with the thresholds if you want, but at least it's disclosed.

The category they didn't disclose is the tell. For requests that looked like frontier AI development, their own system card said the safeguards "will not be visible to the user," and that instead of routing you elsewhere, the system would quietly "limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning." Strip the jargon: we will make the product worse, we decide when, and we won't tell you.

They got caught. After Wired reported it, Anthropic reversed the policy and apologized for "the wrong tradeoff." Good. But don't let the walk-back comfort you. The switch didn't disappear — they just agreed to turn the light on. It came back because of a news cycle, not a principle. And the reversal proves what matters: the vendor's hand was always on the dial.

Now connect it to their own behavior. Anthropic's report, "When AI builds itself," says that more than 80% of the code merged into its codebase is now written by Claude. So the one capability they chose to degrade — AI that builds AI — is the exact one they run at full strength on themselves. They didn't cripple the product. They crippled your copy of it.

Sit in the buyer's chair and feel it. You cannot build a durable business on a capability your vendor can quietly throttle, downgrade, or reserve for a favored list that might include your competitor. When your agent underperforms next quarter, you won't know whether your team shipped a bug or a classifier in San Francisco decided your work looked too ambitious. That uncertainty isn't a side effect. It's the design.

The Ultimate Conflict: Claude Science

Now follow the pattern home. Everything above — Figma, Cursor, the token tax, the switch — is the setup. Here is the payoff for anyone in biopharma. On June 30, Anthropic launched Claude Science, pitching it to pharma as an AI workbench for molecular R&D — sixty-plus scientific capabilities across genomics, structural biology, and cheminformatics. But they didn't stop at building tools for scientists. In the same breath, they announced their own internal drug-discovery program, and demoed the system proposing a candidate for a rare disease on its own.

Think about the architecture of that setup.

They're asking the largest, most profitable pharma companies on earth to trust Claude Science with their most sensitive, multi-billion-dollar R&D pipelines. Meanwhile, Anthropic is running its own pipeline on the same underlying infrastructure. They say they're doing it to "live the challenges alongside the industry." That doesn't resolve the conflict of interest. It defines it. Feed your proprietary targets and molecule structures into Claude, and you're subsidizing a direct, exceptionally well-capitalized competitor.

Anthropic decided that life sciences is too valuable to leave to their customers. Same playbook, every high-value vertical, all at once.

The Path to Architectural Sovereignty

The enterprise is finally waking up to the fact that building your entire business on a single closed-model lab is a strategic dead end. The idea that these labs are neutral utilities is dead.

The only rational defense is model agnosticism. Decouple your application logic and your data from the intelligence layer underneath. Treat models as hot-swappable — because that's what they are.

I'm not the only one saying it. Aravind Srinivas built Perplexity into a multi-billion-dollar company doing exactly this, and he puts it plainly: "The model is not the product. The orchestration is the product; the model is a tool." His logic is that a pure reseller of tokens has no business, because the model itself commoditizes within months — the value lives in the layer around it: the harness, the connectors, the grounding in your real context. More than half of Perplexity's enterprise users already pick more than one model in a single workday. And here's the part the labs can't answer: an agnostic layer can put GPT-5 and Claude side by side in one product. OpenAI and Anthropic structurally can't — you will never find Claude inside ChatGPT. The orchestrator can offer everything; the model owner can only ever offer itself.

Build this way, and you get two things a single-vendor customer never will.

Economic leverage. You route each task to whichever model is cheapest and most capable at that second, and you bypass the metered token trap entirely. The day you can leave is the day the vendor stops setting your terms.

Control over your alpha. You decide where sensitive work runs — including open or self-hosted models for the workloads that matter most — so your proprietary data and workflows aren't concentrated inside the one vendor most able to compete with you. And when a lab throttles a capability, downgrades you to a weaker model, or rewrites its retention terms overnight, you route around it instead of eating it.

If your company's right to exist depends on the hope that a near-trillion-dollar lab won't bother building for your vertical, your thesis is already dead. The companies that survive the next phase won't be the ones with the deepest integration into Claude. They'll be the ones who figured out, early enough, that you can't partner with a vendor who wants your business.