In partnership with
AI labs lose money on cheap tokens as GPU costs surge 82%
TL;DR
AI labs cut token prices to win enterprise customers while GPU rental costs keep rising. One investor's analysis estimates Anthropic's average GPU cost rose 82% in two quarters. A $5 billion short-term compute bill in Q3 flipped an operating profit back into a loss.
Key Takeaways
Labs lock in GPUs on multiyear contracts at $2 to $4 per hour, but spot pricing reaches as high as $13 per hour once reserved capacity runs dry.
A $5 billion short-term compute bill in Q3 turns an operating profit back into a loss, according to the investor analysis.
The top 1% of firms on Ramp spend less on AI per employee in September than in July, which signals fatigue among the heaviest buyers.
Anthropic adds $5.4 billion in revenue in Q3, down from $6.8 billion in Q2.
Adoption across other firms still climbs, even as the biggest spenders stall.
Why It Matters
Many solopreneurs and early-stage startups build their margins on today's low API prices. This story shows that those prices rest on economics that don't hold forever, so token costs could rise once labs need to close the gap between what they charge and what compute costs them.
For founders between pre-seed and Series B, the practical move is to model pricing scenarios now rather than later. Products that keep working at higher token prices, or that switch providers without a rewrite, carry less risk than products that only work at today's rates.
Your GTM Has a Weak Spot. Find It.
Your go-to-market feels stuck. Most of the time, nothing is broken. One or two things are just quietly costing you momentum.
Drop your URL. The free GTM Audit reads your site, pricing, and competitors, scores four dimensions, and hands you a 90-day fix list.
10 minutes. No guessing.
📰 In the News
Headlines & Launches 📣
An X post shows Grok researching the best doctors in a user's area, calling them by phone, and making an appointment. This moves the assistant from answering questions to completing real-world tasks on the user's behalf. For founders building voice agents or booking tools, it signals that general-purpose assistants now compete directly in that space.
According to a report by cybrNK on OSINT Team, bad actors publish fake Claude Code extensions to the Cursor marketplace that install an SSH backdoor on the developer's machine. Developer tools run with elevated permissions, touch source code, and access environment variables, so the attack doesn't need a zero-day. It only needs an install. Small teams without a dedicated security function face the highest exposure here, which makes extension audits a cheap and worthwhile habit.
⚡ Quick Hits
How to run an AI marketing reporting pilot without breaking your data: A checklist tests an AI assistant on one existing report before you connect live systems or hand off distribution.
Stop prompting AI until you've answered these 5 questions: Modi Elnadi of Integrated.Social offers a five-field decision contract that surfaces hidden assumptions before the first prompt.
How did you like today's newsletter?
Thanks for reading,
— Cagri Sarigoz
P.S. Don't Miss Out on These Resources:
🤝 Follow me on X and LinkedIn for regular updates and insights.
