AWS CEO Matt Garman said on the Sources podcast that orbital data centers are unlikely to become commercially viable within 10 to 20 years, citing launch capacity and cadence, vacuum heat rejection at hundreds of megawatts to gigawatt scale, and the difficulty of maintaining and replacing hardware in orbit. He also said AWS is deploying GPUs at one every 14 seconds yet demand still outruns supply, expects capacity shortfalls for two more years, and observes that every 50% cut in inference cost drives order-of-magnitude usage growth.
Inference Cost
推理成本
The computational expense required for a deployed model to process inputs and generate outputs.
Curated Intelligence
Apple is tightening internal AI spending, cutting engineering teams' daily allowance from $300 to $200, with some caps reportedly heading to $100 and non-engineering groups facing limits as low as $25 a day. The move reverses this year's tokenmaxxing enthusiasm and signals a broader corporate reassessment of AI's cost-versus-productivity equation. The report comes from Bloomberg's single sourcing; Apple has not commented.
SpaceX has significantly reduced Grok Bot usage quotas after Elon Musk admitted the operational costs were unsustainable relative to revenue. This move highlights how high inference expenses for AI agents are forcing providers to shift from unlimited access models to stricter resource rationing and tiered pricing.
Data from Anthropic's S-1 filing suggests an ~85% gross margin, implying the internal inference cost for flagship model Opus 5.5 is just $0.60/$3.00 per million tokens. This marginal cost is significantly lower than current market pricing for major models, highlighting the massive efficiency advantages of top-tier labs in large-scale inference.
Amazon CEO Andy Jassy argues that AI inference costs must drop by an order of magnitude to enable widespread enterprise software adoption. With compute capacity booked through 2028, Amazon is leveraging its self-developed Trainium chips to offer 30-40% better cost-performance, aiming to unlock the underserved mid-market and challenge Nvidia's dominance.
MiniMax has officially launched its new lightweight model, M3.1 Flash, achieving nearly 3x faster inference and a 70% cost reduction while maintaining top-tier performance in agentic tasks, coding, and computer use. The model leads in several benchmarks and plans to open-source weights soon, offering developers a high-efficiency solution for complex workflows.
Paper Instruments released Ultramarine 1 Flash, claiming Fable-class performance in finance, law, and consulting knowledge work at 60x lower cost. The model ranks first or second on benchmarks like JobBench and FrontierFinance, with API access now available and weights coming soon.
Anthropic reported $16.8B in Q3 revenue, a 45% quarter-over-quarter increase that exceeds its total 2025 earnings but marks a sharp deceleration from Q2's 140% surge. The slowdown is attributed to seasonal demand dips and competitive pressure from OpenAI's new Astra model, which Anthropic countered with Opus 5.5. As the company prepares for a November IPO, it highlights profitable inference metrics to reassure investors amidst intensifying price wars and closing performance gaps from open-source models.
Simon Willison reports on the launch of GPT 6.1 Sol, claiming it achieves near-Astra intelligence levels at one-fifth the cost. However, the provided text contains only a sponsorship ad, lacking substantive information on technical parameters, benchmark data, or implementation details.
The Berkeley BAIR team highlights that AI inference costs are dropping exponentially, ushering in an era of 'virtually free intelligence.' They argue that as raw intelligence becomes abundant and cheap, the primary bottleneck shifts to building specialized data infrastructure for agents. The article advocates for a paradigm shift towards data systems designed 'for, of, and by' autonomous agents.