Inference Economics: Gartner's 96 Percent

Nvidia's numbers match the scale. Data-center revenue hit $89.0 billion in the quarter ended late July 2026, up 117%². Jensen Huang told investors what changed. "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue." Read the last clause. A cost became a source of income.

A paradox that resolves through volume

Spending nearly doubling looks impossible beside a second trend running at the same time. The price of a token collapsed. One widely cited analysis puts the drop at roughly 95% over two years, and close to 1,000-fold over three³. GPT-4 charged $30 per million input tokens in March 2023. Open-weight models charged about $0.10 by 2026.

Both are true at once, and volume explains it. Companies now run far more inference calls than the price drop alone would predict. Agent work multiplies the calls one job sets off. A support agent handling a ticket in 2023 made one call. An agent system handling the same ticket in 2026 might make a dozen, each one sent to a sub-agent that checks a claim, reads a policy, or drafts a reply. Gartner's dollar figure measures that multiplication. The per-token chart measures only the price underneath it.

Fireworks and Together compete on serving speed

Lin Qiao built Fireworks AI on one bet. Serving open models faster and cheaper than the labs that trained them would win a market of its own⁴. The bet paid in scale this year. A Series D announced July 15, 2026 valued it at $17.5 billion, beside a stated $1 billion in yearly recurring revenue. Qiao led PyTorch at Meta before this, and her founding team came from the same place. Dmytro Dzhulgakov and James Reed both worked on PyTorch core and its compiler. Chenyu Zhao came from leading Google's Vertex AI platform.

Together AI ran a parallel plan under Vipul Ved Prakash⁵. It raised an $800 million Series C in 2026, to press the case that open-source serving beats closed pricing at scale. Ce Zhang serves as chief technology officer. Tri Dao, known for the research behind faster attention, holds the chief scientist title. That gives the company a team built to cut delay and cost out of open-weight serving.

Groq's leadership change and Cerebras' margin warning

Money and a management change arrived together at Groq. The company raised $650 million in June 2026, then $350 million weeks later, reaching a $3.5 billion value⁶. Leadership changed hands between those raises. Adam Winter, who joined in 2024 to run the international business, became chief executive in 2026. Groq's own leadership page confirms the move and skips the reason. Winter now oversees chief financial officer Matt Eng, chief operating officer Alan Rice, and chief technology officer Sinclair Schuller. Their custom chip claims to serve “millions of developers” running “trillions of tokens” weekly.

Cerebras took the opposite path in public. Andrew Feldman, its chief executive and co-founder, rang Nasdaq's opening bell on May 14, 2026⁷. A $5.5 billion raise sent shares up 108% on debut. Five weeks later the stock fell 10%, when Cerebras forecast a thinner margin in its first earnings as a public company⁸. That warning matters past one company. It is public proof that the price collapse reaches the specialists too, even one running its own chips rather than reselling Nvidia parts.

Bounded reasoning attacks the cost side directly

The inference specialists compete on serving efficiency. One research team went after something else: how much reasoning a model has to do per answer⁹. Armağan Amcalar, chief technology officer at OpenServ Labs, and researcher Eyup Cinar published BRAID. It swaps open-ended chain-of-thought for structured, diagram-coded logic. This edition's feature, “Reasoning's Rebate,” covers the paper directly. Its headline result, 74 times the baseline on one test, works the demand side rather than the supply side Fireworks and Together fight over. Amcalar summed up the aim himself. “BRAID boosts performance across every model class, from largest to smallest, making strong reasoning affordable”. If the technique holds past benchmark conditions, Gartner's curve could flatten, because a real share of that $23.3 billion buys reasoning tokens a bounded approach skips¹.

What the spending curve implies for agent infrastructure

Every agent framework, every payment rail for machine-to-machine trade, and every pilot reaching production needs inference cheap enough to run at scale. Model Context Protocol servers and Agent2Agent systems make calls at volumes a human-run workflow rarely reached. That volume is what Gartner measures, while the price per token keeps falling¹.

Buyers reading the 96% figure alone will flinch. Read it beside the token-price collapse, and beside the churn among specialists racing to serve that demand cheaper, and it says something else. A market is still finding its scale, with price and volume pulling against each other, and the outcome open.

By the numbers

• Gartner forecasts $42.276 billion in AI-ready cloud spending for 2026, up 96% from 2025¹.
• Inference takes 55% of that 2026 spending, rising to 59% in Gartner's 2027 forecast¹.
• Data-center revenue at Nvidia hit $89.0 billion for the quarter ended late July 2026, up 117%².
• Inference cost for the same capability fell a reported 95% over the two years into 2026³.
• Fireworks AI reached a $17.5 billion value in July 2026, beside a reported $1 billion in yearly revenue⁴.
• Two rounds, $650 million and then $350 million, carried Groq to a $3.5 billion value in 2026⁶.
• Cerebras shares rose 108% on their Nasdaq debut, then fell 10% weeks later on a margin warning⁷⁸.
• BRAID's bounded-reasoning technique reported a peak performance-per-dollar gain of 74 times on one benchmark⁹.

What to watch

Quarterly earnings from the specialists will show whether Cerebras' margin warning was one company's problem or the category's⁸. A second chipmaker reporting thinner margins would settle it. Gartner's 2027 inference share, forecast at 59%, deserves a check against real spending once the year closes, because agent volume could outrun even that¹. Bounded reasoning reaching production outside benchmark conditions would be the clearest sign that cost, rather than price alone, is finally moving toward companies running agents at scale⁹.

Sources

1. "Gartner Forecasts Worldwide Artificial Intelligence-Optimized IaaS Spending to Grow 96% in 2026," Gartner, Aug. 10, 2026, https://www.gartner.com/en/newsroom/press-releases/2026-08-10-gartner-forecasts-worldwide-artificial-intelligence-optimized-iaas-spending-to-grow-96-percent-in-2026.
2. "NVIDIA Announces Financial Results for Second Quarter Fiscal 2027," Nvidia, Aug. 26, 2026, https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027.
3. "How AI Inference Costs Have Dropped 95% in Two Years, and What Happens Next," Value Add VC, 2026, https://valueaddvc.com/blog/how-ai-inference-costs-have-dropped-95-in-two-years-and-what-happens-next.
4. "Announcing our Series D and $1B ARR," Fireworks AI, July 15, 2026, https://fireworks.ai/blog.
5. "Announcing our $800M Series C to accelerate the shift to open-source AI," Together AI, 2026, https://www.together.ai/blog.
6. Groq leadership and news pages, Groq, 2026, https://groq.com/about-us/.
7. "Cerebras raises $5.5B, then stock pops 108%, in the first huge tech IPO of 2026," TechCrunch, May 14, 2026, https://techcrunch.com/2026/05/14/cerebras-raises-5-5b-kicking-off-2026s-ipo-season-with-a-bang/.
8. "Cerebras falls 10% after chipmaker forecasts shrinking margin in first earnings report since IPO," CNBC, June 23, 2026, https://www.cnbc.com/2026/06/23/cerebras-cbrs-q1-earnings-report-2026.html.
9. Armağan Amcalar and Eyup Cinar, "BRAID: Bounded Reasoning for Autonomous Inference and Decisions," arXiv, Dec. 17, 2025, https://arxiv.org/html/2512.15959v1.