Gemini 3.5 Flash | AI Field Notes #32

AnalysisStarting June 15, Anthropic moves all programmatic Claude usage into its own billing pool, separate from interactive subscription limits. The affected services include the Agent SDK, headless mode (claude -p), Claude Code GitHub Actions, and third-party agent apps. Pro subscribers ($20/month) get $20 in monthly programmatic credits; Max 5x users ($100/month) get $100; Max 20x ($200/month) get $200. Credits expire monthly without rollover. Once the pool empties, additional calls bill at standard API rates. The interactive tier covering web chat, terminal IDE sessions, and Cowork stays unchanged. Anthropic announced this on May 14; users receive an email June 8 to claim credits before the cutover.

AnalysisGoogle launched Antigravity 2.0 at I/O on May 19: a new CLI written in Go (faster and replacing the Gemini CLI outright), a revamped desktop app that can orchestrate multiple parallel subagent workflows, and an SDK for custom agents. Gemini Spark, a persistent agent that runs on dedicated Google Cloud VMs and maintains state between sessions, entered beta for AI Pro and Ultra subscribers. Pricing changed: the existing $250/month AI Ultra plan drops to $200, and a new $100/month tier with 5x the Pro limits was added. Google is asking all current Gemini CLI users to migrate to Antigravity CLI.

AnalysisOpenAI filed a draft IPO prospectus confidentially with the SEC around May 22, with Goldman Sachs and Morgan Stanley advising. The company is targeting a September 2026 public listing at $852 billion to over $1 trillion in valuation, which would rank it between Berkshire Hathaway and Eli Lilly globally. The financial picture is complicated: about $30 billion in annualized revenue, but $14 billion in projected losses for 2026 alone and cumulative losses potentially reaching $44 billion before the company expects to break even around 2029. The S-1, which becomes public roughly 15 days before the roadshow, will also disclose CEO Sam Altman's equity stake for the first time.

AnalysisNvidia's results for the quarter ended April 26 showed $81.6 billion in revenue, up 85% year over year and 20% from the previous quarter. Data Center alone brought in $75.2 billion, up 92% from the prior year. GAAP gross margin held at 74.9%. Nvidia also announced $80 billion in additional share buybacks and increased its quarterly cash dividend from $0.01 to $0.25 per share. Rubin, the next-generation GPU platform that promises 10x lower inference cost per token compared to the current Blackwell line, begins customer shipment in the second half of 2026. AWS, Google Cloud, Microsoft, and Oracle are among the first cloud providers lined up for Rubin-based instances.

AnalysisOpenAI co-founder and former Tesla AI director Andrej Karpathy joined Anthropic's pre-training team on May 19, reporting to team lead Nick Joseph. His stated mandate is to build a new group focused specifically on using Claude to accelerate pre-training research. Pre-training is the large-scale training phase that gives frontier models their baseline knowledge and capabilities; it is also the most expensive and technically demanding part of building a frontier model. Karpathy most recently ran Eureka Labs, an AI-in-education startup. The hire is one of the more significant talent moves in the field this year, and the focus on automating parts of the research process itself signals where Anthropic thinks the leverage is.

AnalysisAnthropic shared financial projections with investors showing $10.9 billion in revenue for the quarter ending June 2026, up 130% from $4.8 billion in Q1. The company expects $559 million in operating income, its first quarterly profit. Compute costs fell from 71 cents per revenue dollar in Q1 to 56 cents, driving the margin improvement. Annualized, the Q2 run rate of $43.6 billion would exceed the annualized revenue OpenAI publicly reports today. Anthropic is targeting an October 2026 IPO at roughly a $900 billion valuation, and the profitability projection gives it a materially stronger pitch to public market investors than its main rival carries into its own roadshow.

AnalysisA federal jury in San Francisco ruled unanimously on May 18 that Elon Musk's lawsuit against Sam Altman, Greg Brockman, OpenAI, and Microsoft was time-barred. The nine-person jury deliberated less than two hours before finding that Musk filed after the statute of limitations expired for claims about harms that arose in 2021 and 2022. Judge Yvonne Gonzalez Rogers agreed and dismissed the case. Musk had already dropped his fraud claims before trial began. He called the result a calendar technicality on X and said he would appeal. The verdict removes the most prominent legal challenge to OpenAI's shift from nonprofit to public benefit corporation, which completed in October 2025.

AnalysisIntuit CEO Sasan Goodarzi announced on May 20 that the company is eliminating roughly 3,000 roles, 17% of its 18,200-person global workforce, to reduce organizational complexity and accelerate AI investment across TurboTax, QuickBooks, and Credit Karma. US employees keep their jobs through July 31 and receive 16 weeks of base pay plus two additional weeks per year of service. The company posted 17% revenue growth and 48% net profit improvement in its most recent quarter, making these cuts a choice rather than a distress move. Intuit has active multi-year partnerships with both Anthropic and OpenAI and plans to embed those companies' models throughout its product suite.

AnalysisColorado's legislature passed SB 189 on May 12, repealing and replacing the original SB 24-205 AI Act before it ever took effect. The original law, scheduled for June 30, 2026, required high-risk AI deployers to maintain duty-of-care obligations, conduct impact assessments, run risk management programs, and report incidents of algorithmic discrimination. SB 189 removes all of those requirements, replacing them with a disclosure-based framework: developers must share documentation with deployers, and deployers must give consumers notices before and after adverse decisions. The replacement takes effect January 1, 2027. A Colorado magistrate had already blocked enforcement of the original law in April while rulemaking was underway.

AnalysisSpaceX's confidential IPO prospectus, portions of which were shared with investors around May 22, disclosed a compute services contract with Anthropic: $1.25 billion per month through May 2029. Annualized, that is $15 billion a year from Anthropic to SpaceX for GPU compute. With Anthropic projecting a $10.9 billion Q2 revenue ($43.6 billion annualized run rate), the SpaceX contract alone represents roughly 34% of its annual revenue run rate. SpaceX is not Anthropic's only compute provider, which makes the figure even more striking as a minimum floor on what frontier AI training and inference actually costs per year.

AnalysisMeta carried out roughly 7,800 job cuts around May 20, completing approximately 10% of its total workforce reduction tied to its AI investment plans. The company committed $65 billion in capital expenditure for AI infrastructure in 2026 and a $135 billion five-year buildout. Cuts concentrated on management and support functions the company described as non-AI-aligned roles, while engineering and AI development headcount was largely preserved. Meta leadership explicitly framed the restructuring as eliminating organizational layers to accelerate product development in AI.