Key Takeaways

  • OpenAI's reported valuation trim from $1.4 trillion to $1.25 trillion stems from relative-value price discovery against public tech comps, not a collapse in buyer demand.
  • Rory O'Driscoll argues that private illiquid companies burning cash while growing at lower rates than top public benchmarks will always price below them.
  • Anthropic saw quarter-over-quarter net new bookings fall after Q2, driven by aggressive token discounting across the industry.
  • Price deflation is hitting model providers directly; Jason Lemkin dropped his own platform expenses by 72% after moving workloads to Haiku 55 and Jev.

The Disagreement

When reports circulated that OpenAI's valuation fell from $1.4 trillion to $1.25 trillion, critics rushed to declare that the private market bubble had popped. Rory O'Driscoll pushed back immediately against the panic.

“I would dispense with the whole AI wave cracking narrative,” (play from 1:22) O'Driscoll said. “I think people are trying to extrapolate their desires, wishes, whatever, prejudice, biases on the story.” (play from 1:24)

For O'Driscoll, the correction is routine balance sheet math. Private assets trade against liquid public stocks. If public megacaps trade at specific multiples while generating billions in free cash flow, an illiquid private lab cannot float untethered from reality. As O'Driscoll framed the pricing logic: “What would you pay for the smaller company growing less quickly, still losing money that's illquid? Answer less than 1.5 trillion. How much less? Well, that's what they're figuring out.” (play from 1:40)

Jason Lemkin, however, tracked the issue to the revenue line rather than market sentiment. Multiple compression tells only part of the story. Aggressive price cuts are eating directly into top-line revenue.

“Anthropic had a down quarter from net new revenue,” (play from 2:28) Lemkin noted. “It was down in this last quarter in terms of net new bookings over Q2.” (play from 2:30)

Lemkin argued that institutional backers will panic when growth flatlines. “I think folks are just going to be nervous about a deceleration in quarter-over-quarter growth. And I think they're going to be nervous over pricing pressure.” (play from 7:14)

To illustrate how steep that pricing pressure is, Lemkin pointed to his own bills. After switching workloads to cheaper alternatives, his costs fell off a cliff: “When I switched it myself to Haiku 55 and Jev, total costs went down about 70%. 72%.” (play from 14:09)

Who's Right (and When They're Wrong)

Both investors are identifying halves of the same economic vice.

O'Driscoll is right about financial gravity. Private valuations do not exist in a vacuum. When private market investors demand discounts for illiquidity and ongoing operating losses, they are performing basic math. But O'Driscoll misses the operating threat: if top-line revenue growth stalls, relative-value adjustments turn into permanent valuation resets.

Lemkin is right about the commercial squeeze. When model providers slash token prices to protect market share, they create severe top-line deflation. If your customers consume 50% more tokens but pay 72% less per token, your net new revenue contracts even as product usage climbs. Anthropic's sequential dip in net bookings proves that volume gains are not yet outpacing price deflation.

Where Lemkin's warning overstates the danger is treating gross revenue deceleration as a collapse in commercial adoption. Lemkin's 72% cost savings on Haiku 55 did not reduce his team's dependency on Anthropic; it entrenched it. Lower token costs expand the addressable workload for application builders, even if they squeeze foundation model revenue in the immediate quarter.

What to Do With This

Audit your model spending across your entire stack before Friday. If your engineering team runs standard queries on flagship models, migrate high-volume background tasks to lighter models like Haiku 55. Target the 70% cost reduction Lemkin achieved, then route those infrastructure savings into product distribution rather than foundation model compute.