Today's News
OpenAI via Exponential View
OpenAI has completed Jalapeño, an inference-only ASIC designed in partnership with Broadcom, achieving industry-leading performance in just 16 months—a compressed timeline enabled by using frontier models to write kernels and optimize compute blocks. The chip beats every tested Nvidia, AMD, and Google competitor on tokens per megawatt without speculative decoding, hitting over 700 tokens/sec/user at low concurrency on DeepSeek R1 and ~1,400 tok/sec/user on other models, while matching Nvidia's accuracy on GSM8k evals.
Jalapeño's architecture prioritizes power efficiency over specialization, using HBM4 and avoiding prefill-decode disaggregation to trade theoretical efficiency for operational pragmatism. On perf/TCO, it matches Vera Rubin's output tokens per dollar, though Jalapeño achieves this without speculative decoding—meaning future implementation of that technique could extend its cost advantage further. The results validate the emerging heterogeneous chip market where specialized architectures will fragment compute rather than consolidate it, directly challenging Nvidia's inference dominance and signaling that frontier labs can compress hardware design cycles through AI-assisted optimization.
Nvidia projects $700B annual revenue while doubling supply commitments to $279B, crystallizing its role as global AI deployment bottleneck
Limitless
Nvidia's scale has reached inflection: at $700B projected annual revenue, each of its 42,000 employees generates revenue equivalent to a mid-size American business. The company has doubled its supply commitments to $279B to lock up manufacturing capacity, underscoring the intensity of competition for foundry slots. Meanwhile, the top five cloud companies plan to spend $1.3 trillion on AI infrastructure buildout next year—more than Spain's GDP—making Nvidia's ability to deliver chips the critical constraint on global AI deployment speed.
Open-weight models hit 62% token share at Vercel; fine-tuned versions match frontier models on specific tasks at 1/14th the cost
Azeem Azhar, Exponential View
Open-weight models have captured 62% of token share at Vercel, with fine-tuned variants now matching or beating frontier models on specific tasks. Bridgewater fine-tuned Qwen to achieve 30% fewer errors than closed models at one-fourteenth the inference cost; GLM-5.3, released this week, reshapes the cost-performance Pareto frontier. This signals meaningful competitive pressure on frontier model providers' inference revenue, though frontier models retain advantages in service guarantees and general capability.
Lenny's Newsletter
OpenAI's head of Codex and ChatGPT Work, Tara Seshan, frames the emerging product paradigm as a shift from "rowing" (execution) to "steering" (judgment), with human ambition becoming the differentiator as AI handles execution. OpenAI builds products for model capabilities 2–3 months ahead of current state rather than optimizing for today's models, reflecting how frontier labs think about product-market fit in a fast-moving capability environment. The strategy emphasizes that ambition itself has become the bottleneck for companies, making the PM role fundamentally about elevating others' ambitions.
Zumper via GrowSF
San Francisco's median one-bedroom rent has reached an all-time high of $4,300, up 25.7% year-over-year and now just $200 behind New York City's $4,500—the narrowest gap since February 2022. The surge is driven by AI-driven hiring colliding with a construction pipeline that remains thin; apartment vacancy has fallen to 3.7% citywide, the lowest level since 2019. Developers signal that another 15–20% rent increase could trigger 12–24 months of construction after the city reduced development impact fees by 67% and lowered inclusionary zoning requirements to 5%, suggesting policy changes may finally unlock supply response.