Moving Up the AI Stack
- The AI Labs, side by side — $65B vs $40B run rates, two IPO clocks, and a new contest over trust.
- AI Applications — Coding, one of the largest AI applications is approaching $10B+ run-rate
- What is driving growth in applications? — the cost of intelligence is rapidly declining
For most of this year, we’ve followed the cash down the stack into chips, memory, networking and power.
This week, we’re focusing in on the other layers of the AI stack: models and applications.
1. The Labs, Side by Side
Two frontier labs, two revenue prints, two different playbooks.
- . Revenue run rate reportedly surpassed $65B at the end of July, up from ~$47B mid-year and ~$9B at the end of 2025. Investors are reportedly targeting a ~$2 trillion valuation in an October IPO, the largest ever, eclipsing , with run rate expected to hit $100B to $120B by year-end.
. Run rate recently passed $40B, up ~35% quarter-to-date, with enterprise up ~50%. CFO Sarah Friar told staff it expects to be public in 2027, sooner if growth inflects, and noted Anthropic could reach the public markets first.
OpenAI also flagged a mix shift most missed: business revenue now outweighs consumer, reversing an earlier 60/40 tilt.
And the competition is opening a front that isn’t about benchmarks at all: TRUST:
- OpenAI unveiled Private Safety Processing, which monitors for misuse while retaining none of a customer’s data aimed squarely at Anthropic’s retention practices.
- Anthropic, for its part, is reportedly revising its own data-retention policy.
The thread: as capability converges, the contest is shifting toward what enterprises will let these systems touch.
2. AI Applications: Coding
Coding has emerged as one of AI’s largest and fastest-growing application categories, already representing an estimated $10B+ market.
Cursor set tech industry records by scaling from launch to over $100 million ARR in about 12 months, surging past $1 billion ARR by late 2025, and exceeding $3 to $4 billion in annualized revenue by mid-2026.
SpaceX closed its ~$60B acquisition of Anysphere, Cursor’s parent, the largest startup acquisition on record, folding it into a new SpaceXAI unit, implying approximately 15-20X Revenue multiple.
- Days later, Cursor launched Origin, a code-hosting platform aimed directly at GitHub. The timing had an edge: it shipped in the middle of a major GitHub outage.
That second move is the tell: a code editor expanding into hosting and, potentially, the rest of the software lifecycle.
And it isn’t alone:
- OpenAI. Its coding product has reached 20 million weekly users.
- Anthropic. Claude Code has become a signature of its enterprise and developer traction.
Stripe. Reportedly agreed to buy — the layer that routes model calls for more than $7B, about 5x its valuation three months earlier.
3. The Cost of Intelligence
Underneath the revenue and the deals, the price of intelligence is falling fast.
LLM Expenditure Index (price per million tokens) peaked near 2.05 in early June and sits around 1.02 as of August 19. Roughly a 50% decline, essentially back to where it started in December.
The intuitive read is commoditization. The data underneath says something more interesting:
- BCG studied 107 public tech companies (>$500M revenue): the heaviest token users grew revenue 16.5% year-over-year at the median, versus 5.1% for the lightest, measured through coding environments like Cursor.
- Higgsfield, an AI-content app, just disclosed $700M in annualized revenue, with agentic-product users up 42x in three months.
- Google is now processing ~3.2 quadrillion tokens a month, about 7x a year ago. Projections for token demand are expected to be up 24x by 2030.
Falling prices don’t look like softening demand. They look like intelligence getting cheap enough to pour into far more places.
But the floor is rising too. Cheap open-weight models, China’s GLM among them, now land within reach of Claude and GPT on coding at a fraction of the price, with some enterprises citing 60 to 90% savings. That gives buyers a real off-ramp from frontier pricing, and it is the crux of the “who captures the value” debate.
Even so, frontier tokens still appear to capture roughly 65 to 85% of the economic value, while open source takes around 80% of the volume. We believe there is a place for frontier models too.
The thread: the unit price is dropping, but consumption, and the revenue built on it, is climbing faster. Cheaper tokens are expanding the pie.
What We’re Watching
Questions, not answers, but the ones this week put in front of us:
- Does owning the workflow become durable economics? Coding, content, and agents own where the work happens. Whether that becomes lasting margin, or just distribution someone else monetizes, will dictate where the value accrues.
- As tokens get cheaper, who keeps the surplus? A ~50% drop in price has to land somewhere: with the labs, the applications, or the end customer.
We’ve spent years tracing AI’s dollars through the physical layers of the buildout. This week, we moved up the stack to see where those trillions of dollars in infrastructure investment may ultimately get monetized.
The application layer is still small relative to the trillions expected to flow into AI hardware and infrastructure. But the numbers are becoming meaningful, the business models are taking shape, and the competitive landscape is moving quickly.
It’s still early, but it’s time to start mapping where the value will accrue.