The Open-Weight Squeeze on Frontier AI

When Product-Market Fit Backfires

In 2007, Marc Andreessen wrote an essay that became the founding gospel of modern venture capital: for an early-stage founder, the only thing that matters is getting to product-market fit.

In Andreessen's formulation, once the market starts aggressively pulling a software product out of a startup, the founders can screw up almost everything else—sales, hiring, management—and still build an enormously successful company.

But agentic AI has introduced an awkward complication to that formula: what happens when the cost of satisfying customer demand grows faster than the revenue it generates?

Every time an AI agent reasons through a problem, browses a webpage, or retries a failed task, it consumes computational tokens. Those tokens carry an ongoing cost, often billed directly by OpenAI, Anthropic, or another model provider.

For startups charging fixed subscription fees, that creates a potentially dangerous mismatch between revenue and costs. Customer usage can grow far faster than the money coming in.

In other words, a startup can experience precisely the kind of explosive customer adoption that venture capitalists celebrate—and find itself heading toward a financial disaster.

The $1.50 Problem

Consider Harvey, the $15.6 billion legal tech startup backed by Sequoia Capital and widely viewed as the gold standard of enterprise AI adoption.

As Bloomberg reported last month, Harvey experienced a brutal reversal in its financial economics just as demand was surging:

After a March update to its AI agents, Harvey's customer usage spiked, but its gross margins dropped sharply—from about 50% at the beginning of the year to -50% by June, according to a person familiar with the matter.

To appreciate the absurdity of that shift, you have to run the math. A software company with a negative 50% gross margin is in the business of paying $1.50 in raw compute costs to deliver every $1.00 of revenue it collects from its clients. The more successful the product becomes, the faster the startup incinerates its cash reserves.

The culprit was a twenty-fold surge in token consumption over the course of the year, which Harvey's subscription pricing simply couldn't cover. And the problem extended well beyond AI startups. Uber reportedly exhausted its entire annual 2026 AI budget by April after encouraging its software developers to "tokenmaxx" on Anthropic's Claude Code.

For Harvey, the answer was to reduce its dependence on expensive proprietary models. In August, it released an in-house model built on Moonshot AI's Kimi K3—a Chinese open-weight model that promised comparable performance at a fraction of the cost. The move helped bring Harvey's gross margins back into positive territory.

The 45-Billion-Token Calculation

This margin panic is not confined to tech companies. It has triggered an enterprise-wide flight to margin safety across the Fortune 500.

As the Financial Times recently reported, corporate executive mentions of "open weight" or "open source" AI models on earnings calls surged sixfold in August and September compared to the same period last year.

The shift is already visible in the infrastructure data. According to an analysis by Social Capital, in the final week of September, DeepSeek models alone processed more tokens through OpenRouter than OpenAI, Google, Anthropic and xAI models combined.

The reason is simple: at true enterprise volume, almost nobody can afford frontier pricing.

Here's Andy Markus, chief data and AI officer at AT&T, speaking with the FT:

"We're doing 45 billion tokens a day," Markus said. "At that scale, costs become super important. If we can move that traffic from a super expensive closed-source model to an open model that we can often serve at [a fraction] of the cost, and still maintain the accuracy, then we'll do that."

Markus said AT&T took open models and tuned them using the company's proprietary data so they performed as well as, or better than, closed alternatives for specific tasks. "The open models are getting better and better," he said. "That's giving us more optionality."

AT&T now runs 40% of its massive AI traffic on open-weight models and aims to push that to 70% within the next twelve months.

If an open-weight model can handle a customer service query, summarize an internal document, or extract information from a database, why pay a premium to send that work to OpenAI or Anthropic?

The Price of Intelligence

For the frontier labs, this creates an uncomfortable problem.

OpenAI is reportedly seeking private funding at a $1.2 trillion valuation, while Anthropic is preparing for an IPO targeting $2 trillion. Both are betting that demand for AI intelligence will continue to grow at an extraordinary pace.

There is good reason to believe it will. On a recent Bloomberg podcast, OpenAI CFO Sarah Friar described how an 80% price cut for its lightweight Mini models was followed by a tenfold increase in token usage.

That is an extraordinary response. It suggests that cheaper intelligence can unlock enormous amounts of additional demand.

But growing demand and capturing its economic value are two different things. As customers learn to use cheaper models for routine work, the frontier labs may find themselves processing vastly more requests while facing relentless pressure to reduce the price of each one.

Harvey discovered the pressure from one side of the transaction: customers were using more AI than its subscription fees could support. AT&T is approaching it from the other side: if cheaper models can do the job, why keep paying the premium?

And that may be the hardest commercial challenge facing the frontier labs. They can build increasingly powerful models, watch AI adoption explode, and still struggle to capture as much of the resulting economic value as their investors expect.

Signal Stack

The operating reality beneath the headlines.

• Bloomberg Doubles Down on Conversational AI (WSJ) – Bloomberg built its AI on 9,000 engineers, proprietary market data, certified sell-side research and its own chat network rather than the open web—a reminder that in regulated professional markets the constraint on AI products is trusted data and distribution, not model capability.
• LSEG Taps Nvidia for Faster Historical Market Data Processing (Reuters) – A stock exchange operator is now an Nvidia customer, which is what the demand story beyond frontier labs and hyperscalers actually looks like: incumbents applying accelerated compute to datasets they already own rather than training anything.

You received this message because you are subscribed to ARPU newsletter - you may unsubscribe anytime at the bottom of this email.

If a friend forwarded you this message, sign up here to get it in your inbox.