AI Tooling Cost Squeeze | AI Field Notes #12

AnalysisGitHub announced on April 27 that all paid Copilot plans (Pro, Pro+, Business, Enterprise) will move to usage-based billing on June 1, 2026, replacing the old request quota with GitHub AI Credits priced on tokens in, tokens out, and cached tokens. Base prices stay flat, $10 to $39 a month, but a question and a multi-hour agent run will no longer cost the same. GitHub said the weekly cost of running Copilot has nearly doubled since January, and signups for Pro, Pro+, and Student were paused on April 20 to stop trial abuse. Annual subscribers face higher model multipliers. Copilot code review will also burn GitHub Actions minutes. The era of all-you-can-eat agentic coding is closing, and the meter is the new ceiling on how aggressive your AI workflow can be.

AnalysisOn April 29, Sam Altman said OpenAI will start rolling out GPT-5.5-Cyber, a security-tuned variant of its newest flagship, to critical cyber defenders within days through the Trusted Access for Cyber program. The base GPT-5.5 model carries a High rating on OpenAI's internal cybersecurity risk scale, one notch below the Critical threshold for autonomous zero-day exploit creation. Unlike Anthropic's Claude Mythos, which was capped at roughly 50 organizations after it was found capable of chaining vulnerabilities autonomously, OpenAI plans broader distribution: governments, critical infrastructure operators, security vendors, cloud platforms, and financial institutions. No technical specs or benchmarks have been published. Two of the largest AI labs now treat their top cyber models as controlled exports. The frontier just split between what ships to ChatGPT and what ships only to people with badges.

AnalysisBloomberg reported on April 29 that Anthropic is considering preemptive offers of roughly $50 billion at a valuation between $850 billion and $900 billion, more than double its $380 billion mark from February. The board is expected to decide in May. Run-rate revenue has gone from about $9 billion at the end of 2025 to over $30 billion in April 2026, with internal estimates closer to $40 billion driven largely by Claude Code. Customers spending more than $1 million a year now exceed 1,000, double February's figure. The number under negotiation would edge past OpenAI's $852 billion March round, just as OpenAI quietly missed internal targets for revenue and weekly users. Two years of OpenAI being the unquestioned center of the AI market is over.

AnalysisMicrosoft reported on April 29 that Azure cloud revenue grew 40% in the fiscal third quarter, with its AI business now above a $37 billion annual run rate. Total revenue was $82.9 billion. CFO Amy Hood said Q4 capex will exceed $40 billion and full-year 2026 spend should land around $190 billion, with about $25 billion of the increase coming from higher component prices. CEO Satya Nadella said Azure will stay capacity constrained through 2026, meaning the bottleneck is not demand. Microsoft also restructured its OpenAI deal so OpenAI can now serve products on any cloud while Microsoft keeps a non-exclusive license through 2032 and a capped revenue share. The cloud advantage has shifted from owning the lab to owning the racks the lab needs.

AnalysisThe Financial Times reported on April 30 that SoftBank is preparing a US-listed spinout called Roze, targeting a valuation around $100 billion and an IPO as soon as this year. Roze would bundle ABB Robotics (which SoftBank agreed to buy last year) with energy, land, and infrastructure assets, and use autonomous robots to speed data center construction. Masayoshi Son is driving it personally and an analyst day is planned for July at a Texas site. The pitch is honest about the bottleneck: with hyperscaler capex on track to top $600 billion in 2026 and a US shortage of skilled trades like electricians and HVAC technicians estimated above 400,000 workers, chips and power are no longer the only choke points. SoftBank also needs the cash to cover more than $30 billion in remaining OpenAI commitments.

AnalysisThe Department of Justice intervened on April 24 in xAI's lawsuit against Colorado's AI Act, the first time the federal government has moved to invalidate a state AI law. Four days later, a federal magistrate judge granted a joint motion to stay enforcement, which was set to begin June 30. The Colorado law is the most comprehensive state AI regime in the country, with risk management, impact assessments, and disclosures for high-risk systems used in hiring, housing, lending, education, and healthcare. The DOJ is acting under President Trump's December 2025 executive order, which created an AI Litigation Task Force specifically to challenge state AI rules. Other state laws in California, Texas, and Illinois are next on the list. The federal government is no longer the backstop for state AI rules. It is the prosecution.

AnalysisAlphabet reported on April 29 that Google Cloud revenue grew 63% year over year to $20 billion, more than double last quarter's growth rate. CFO Anat Ashkenazi said the cloud backlog nearly doubled in a single quarter to $462 billion, with about half expected to convert to revenue in the next 24 months. Capex guidance went up to $180 billion to $190 billion for 2026, and Ashkenazi said 2027 will be significantly higher. CEO Sundar Pichai said Gemini Enterprise paid monthly active users grew 40% quarter over quarter and revenue from products built on Google's GenAI models grew nearly 800% year over year. Alphabet shares jumped 8% in after-hours trading. The Anthropic deal, $40 billion in cash and 5 gigawatts of TPU capacity, looks less like a bet now and more like the trade that keeps printing.

AnalysisThe Washington Post reported on April 28 that Fermi America, the startup behind the Donald J. Trump Advanced Energy and Intelligence Campus near Amarillo, Texas, is in trouble after its stock crashed. Co-founded by former US Energy Secretary Rick Perry, Fermi went public in October 2025 with plans to build a nuclear-powered AI data center campus. Eight months later, the financing math is wobbling and the timeline for nuclear capacity is doing what nuclear timelines tend to do. Meanwhile Pure DC has paused its Middle East AI infrastructure investments after an Iranian attack hit its Abu Dhabi data center, and Maine just vetoed a data center moratorium on April 29. The story underneath: the AI buildout assumes power and political stability that not every site can deliver. The cap on growth is no longer chips. It is the grid and the geopolitics around it.

AnalysisAmazon's Q1 2026 earnings on April 29 included a number that stood out from the cloud noise: Bedrock, AWS's hosted model service, processed more tokens in Q1 alone than in every prior quarter combined, with customer spend up 170% quarter over quarter. AWS revenue grew 28% and shares finished April 30 at $265.06. The chip story matters here too: Anthropic, the largest customer on Bedrock, is now committed to spending more than $100 billion on AWS technologies over the next 10 years, with Trainium chips taking a larger share. The number reframes the AI adoption question. The stat people quote about ChatGPT is users. The stat that says where serious work is happening is tokens, and on Bedrock that curve just went vertical.

AnalysisReuters reported in late April that Meta is installing software on US employee work computers that captures mouse movements, clicks, keystrokes, and selected screenshots, then feeds that data into the company's AI training pipeline for Meta Superintelligence Labs. A staff memo framed the rollout as a way for rank-and-file workers to help models learn basic computer-use behavior, like navigating dropdown menus and using keyboard shortcuts, by simply doing their daily jobs. The same company is preparing to lay off 8,000 staff starting May 20, then asking the survivors to be the training set for the agents that will replace another wave. Meta acquired a 49% stake in Scale AI for more than $14 billion last year, and former Scale CEO Alexandr Wang now leads the lab. The data labeling problem is being solved at the desk.