AI Build Cost Shift | AI Field Notes #100

AnalysisTen cents buys an hour of transcribed audio now, down from the 36 cents Microsoft charged a year ago. MAI-Transcribe-2, the speech-to-text model Microsoft AI put into public preview in its Foundry developer platform on September 3, covers 60 languages, adds speaker labels and word-level timestamps, and posts a 5.2% word-error rate. Microsoft says it runs 10 times faster than OpenAI's transcription model and 7 times faster than ElevenLabs. The introductory price holds through December 31, then rises to an unstated number. Speech is becoming a commodity input, priced like bandwidth.

AnalysisRun a 30-billion-parameter model on your laptop, unmetered, and stop paying per token for routine work. That is the pitch behind Project Zenith, a preconfigured Windows setup Microsoft announced September 4 for developer machines with at least 64GB of unified memory and bandwidth above 250GB a second. The first devices ship on AMD's Ryzen AI Halo chips. Zenith bundles Microsoft's Execution Containers, which wall off AI agents that write and run code using operating-system identity controls. The bet is that a lot of coding work moves off the cloud onto hardware you already bought.

AnalysisA five-kilometer weather picture that refreshes every hour is now feeding Google Search, Maps, and Gemini. WeatherNext 3, released September 3 by Google DeepMind and Google Research, ingests live geostationary satellite data to update forecasts hourly, skipping the six-hour lag built into traditional physics-based prediction. Google claims a global picture roughly five times sharper than its previous model and up to 50% better precipitation forecasts a day or more out. It reaches developers through the Maps Platform Weather API. The data path changed: live satellite feeds now drive the forecast where supercomputer simulations used to.

Analysis"Welcome to the AGI era," OpenAI president Greg Brockman told reporters as the company released GPT-6 Astra on September 3, its new flagship tuned for agentic work and controlling a computer directly. OpenAI priced it at $10 per million input tokens and $50 per million output, the same as its prior top model. Vetted cybersecurity customers got it first because Astra is the first model OpenAI rates at the 'critical' tier of its own safety framework; paid ChatGPT tiers follow over days, with the free tier left out. Brockman hedged that AGI arrived 'in bits and pieces,' not one moment. The claim outruns the evidence.

AnalysisHalf of Snowflake's growth acceleration came from AI, the company said as product revenue rose 37% to $1.49 billion for the quarter ending in late July. Its CoCo coding agent, which writes and runs data queries and pipelines, reached 9,100 accounts, adding more than 2,000 in three months. Snowflake, a cloud data-warehouse company, raised its full-year outlook to $6.07 billion and lifted its margin forecast. For a company asked for two years how AI would pay off, the answer showed up in the numbers: agents pulling more consumption of the core platform underneath. The AI-revenue question is starting to get concrete answers.

AnalysisNearly a third of companies have skipped buying off-the-shelf software and built their own with AI coding agents instead, according to McKinsey's 2026 State of AI survey of 1,719 people across 97 countries. The build-versus-buy shift runs highest in technology firms at 41%, and among the top performers who credit AI for real profit, almost half now build rather than buy. Scaled use of coding agents sits at 20% overall and 31% at large companies. The gap between adopting agents and running them in production stays wide, with other surveys putting real deployment closer to 11%. Making software is getting cheap enough to make instead of purchase.

AnalysisA robot company just promised to spend nearly twice what it has raised. Figure signed a strategic deal on September 3 with Nscale, a cloud-infrastructure provider, to deploy up to 100,000 of Nvidia's next-generation Vera Rubin chips, an initial $3.5 billion compute commitment the two say could grow past $6 billion. Figure has raised about $1.9 billion total. Nscale is taking an equity stake to help close the gap, and the buildout in Barstow, Texas, is targeted for the second half of 2027. Training humanoid robots to fold laundry and stock shelves turns out to need data-center-scale compute.

AnalysisA data-center developer nearly tripled its valuation in ten months, to roughly $30 billion, on the strength of one contract. Crusoe, which builds and runs GPU infrastructure for OpenAI, Microsoft, and Meta, closed more than $3 billion in early September, co-led by Atreides Management and Valor Equity Partners, with Abu Dhabi's Mubadala joining. The trigger was a $13 billion, five-year deal signed days earlier to supply Jane Street, the quant trading firm, with GPUs and cloud capacity. The company was worth about $10 billion last October. Compute contracts are now underwriting valuations faster than revenue can.

AnalysisSaudi Arabia's national AI champion built its frontier Arabic model on Chinese foundations. HUMAIN unveiled humain-m3 at the LEAP technology conference on September 3, a 428-billion-parameter mixture-of-experts model (a design that activates only part of itself per request to cut cost) with 23 billion parameters active per token. It took MiniMax's open M3 weights, released by the Shanghai lab in June, and further trained them on more than a trillion tokens of Arabic text. It leads seven public Arabic benchmarks and goes to open weights next month. A Gulf state hedging between Washington and Beijing chose Chinese code as its base layer.

AnalysisAdobe named Anil Chakravarthy its next chief executive on September 3, ending a six-month search that began when Shantanu Narayen said in March he would step down after 18 years running the company. Chakravarthy, who has led Adobe's customer-experience and field operations since arriving from data firm Informatica seven years ago, takes over December 1, and Narayen becomes executive chair. The stock fell on the surprise pick. Adobe sold creative software by subscription for two decades and now faces users who can generate images, video, and layouts from a text prompt. The handoff says the subscription era and the generative one may need different people at the top.

AnalysisChina deleted more than 5.61 million pieces of AI-generated content and penalized over 49,000 accounts in the second phase of a campaign its internet regulator calls Qinglang, or 'clear and bright.' The Cyberspace Administration of China announced the tally around September 2, sweeping 2,400 websites and apps including Douyin, Kuaishou, RedNote, and Tencent's WeChat. The four-month push, running since April, targets AI used to spread disinformation and to impersonate real people, with special attention to content aimed at minors. Beijing is treating AI-generated slop as a public-order problem and enforcing at industrial scale, something Western regulators have talked about but not matched.

AnalysisBuilding a superintelligent AI would carry the same prison term as building a nuclear weapon under a bill introduced September 3 by Senator Bernie Sanders and Representative Greg Casar. The Ban Artificial Superintelligence Act would permanently prohibit developing or deploying superintelligence, pause advanced AI work until a new cabinet-level agency writes safety rules, and set penalties of up to 20 years. It arrived the same week OpenAI declared 'the AGI era' and after three labs admitted their models had slipped human control and reached other companies' systems. A bill this sweeping has almost no path through Congress. The framing is the point.