AI Cost Reality Check | AI Field Notes #3

AnalysisBloomberg and the Taipei Times reported this week that financial regulators across Asia, led by the Monetary Authority of Singapore, are pushing banks to close security gaps after Anthropic (the AI lab behind Claude) disclosed that its unreleased Mythos model found thousands of previously unknown software bugs during internal testing. Anthropic is keeping Mythos inside a 50-company safety program it calls Project Glasswing, but public knowledge that one frontier model can surface zero-days (fresh, unpatched exploits) at scale has already raised the regulatory floor. PYMNTS reported this week that AI is now involved in roughly 76% of cyberattacks against banks. The shift worth naming: a capability that is not publicly deployed is already setting the floor for compliance. Regulators are treating Anthropic's caution as a grace period.

AnalysisOn Monday, reports surfaced that Google is in talks with Marvell (a chip-design firm that is not a household name) to co-develop a memory processing unit paired with Google's TPU, the custom AI chip family Google designs in-house as an alternative to Nvidia GPUs. Marvell's stock popped, Broadcom fell roughly 3% because it already holds much of Google's TPU manufacturing business. The subtext: Google no longer wants a single vendor controlling its custom-silicon roadmap, and every other hyperscaler (a giant cloud provider like AWS or Azure) is arriving at the same conclusion. Nvidia still sells the fastest training chip, but the real 2026 money is in inference (running the model for paying users, as distinct from training it), and Google wants cheaper specialized silicon for that workload. Single-vendor leverage has a shelf life.

AnalysisAnthropic made Claude Opus 4.7 generally available on April 16 across its own API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft's Foundry, with GitHub Copilot adding it the same day. Reported scores include 87.6% on SWE-bench Verified (a benchmark of real open-source software bugs), up from around 81% on the prior Opus. Pricing held at $5 per million input tokens and $25 per million output tokens. The interesting choice is the removal of explicit developer-set "thinking budgets" in favor of adaptive reasoning, meaning the model decides how hard to think on its own. That design change matters more than the headline benchmark. It concedes that human-set reasoning budgets were always guesses, and Anthropic is willing to make spend less predictable for developers in exchange for better long-running agent performance. Other labs will follow.

AnalysisMaine's legislature passed, and its governor is set to sign, the country's first statewide moratorium on new data centers larger than 20 megawatts, roughly the footprint of a mid-size AI training facility. The freeze runs until late 2027 while state energy and environmental regulators draft rules. Context: Data Center Watch tracked about $156 billion of data-center projects cancelled or delayed across the US in 2025 in the face of local opposition, and Stanford's 2026 AI Index puts global AI data-center draw at 29.6 gigawatts, enough to power New York State at peak. Maine is small. It is also a template other states are already copying. The fight is familiar from earlier industrial cycles. A local community asks who gets the jobs and who gets the electric bill, and the answers stop being automatic.

AnalysisVercel (a widely used hosting platform for web apps) disclosed on April 19 that attackers reached its internal systems through a compromised AI tool called Context.ai that one employee had authorized with a Google Workspace OAuth token, the familiar "sign in with Google" permission that also grants mail and file access. Customer data was stolen and is now listed for sale on a cybercriminal forum, with the seller claiming ties to ShinyHunters, a known extortion crew. The pattern is hardening: an AI productivity plugin sits outside normal security review, carries broad permissions, and becomes the soft entry. Mercor (an AI-data startup valued at $10 billion) was hit the same way earlier in April through another developer tool. The AI developer stack is quietly the fastest-moving unreviewed software supply chain inside most companies.

AnalysisOver the weekend in Beijing, a humanoid robot from a Chinese startup finished a half-marathon course ahead of most human participants, with several outlets reporting it edged past a reference human half-marathon time. The race is a stunt, the biomechanics are carefully staged, and the robot probably cost more than a small apartment. The unglamorous takeaway: China's humanoid ecosystem is now shipping hardware that can operate in uncontrolled outdoor environments for hours at a time. The same day, Shenzhen-based X Square Robot closed a Series B of about $276 million led by ByteDance (the Chinese parent of TikTok) and Sequoia China to build general-purpose robots for industrial work. Stanford's 2026 AI Index already notes China leads in robotics patents and publications. The private market is catching up to that finding.

AnalysisOn April 4, Anthropic cut off several third-party agentic tools, including OpenClaw (an open-source wrapper that runs coding agents autonomously), that were running nonstop AI sessions on $200-per-month Max subscriptions and extracting what would have cost up to $5,000 in pay-per-token usage. The company is now steering enterprise deals toward per-token billing, where customers pay for actual compute used rather than a flat monthly fee. OpenAI is holding flat-rate plans, and ChatGPT head Nick Turley conceded on a podcast this month that "unlimited" AI may work about as well as unlimited electricity. Finance tool Ramp reports AI costs across its customer base are up roughly 13x in a year and nobody knows how to budget. Both companies are aiming at IPOs this year. The one that priced for reality will have cleaner revenue to defend.

AnalysisX Square Robot, a four-year-old Shenzhen startup building general-purpose humanoid and manipulation robots, closed a Series B of roughly 2 billion RMB, about $276 million, announced on April 20. ByteDance and Sequoia China led. The company is already shipping units called Quantum-1 and Quantum-2 into industrial-manufacturing and logistics pilots. Paired with the Beijing half-marathon robot result the same weekend, the shape of the 2026 robotics funding cycle becomes clearer. Chinese firms are raising capital on shipped hardware while US humanoid startups are still raising mostly on demo videos. Stanford's AI Index already flagged that China leads in robotics patents and publications; private capital is now agreeing. This is the kind of story that looks like one round today and a supply-chain fact in two years.

AnalysisSnap CEO Evan Spiegel told staff on April 15 that the company will cut about 1,000 employees, 16% of its workforce, and close another 300 roles before they were filled. Spiegel attributed the decision to AI advancements that "allow smaller teams to achieve the same output," and disclosed that AI now writes more than 65% of Snap's new code. The company expects about $500 million in annualized savings and the stock rose 11% in premarket. The unusual part is the specificity. A named percentage, a named savings number, a named pressure from activist investor Irenic Capital, which had publicly demanded cuts of up to 21%. Most AI-linked layoffs hide inside phrases like "optimizing for future innovation." Spiegel dropped the euphemism. Other CEOs will notice that the stock responded well to the direct version.

AnalysisThe US House Select Committee on the Chinese Communist Party released an investigation this week documenting specific pathways by which restricted Nvidia AI chips reached Chinese buyers despite export controls: Southeast Asian shell companies, resale through cloud brokers, and distillation, a technique in which a smaller model is trained on the outputs of a larger one to clone capability without copying weights. OpenAI, Anthropic, and Google confirmed earlier this month via Bloomberg that they are coordinating through the Frontier Model Forum (an industry nonprofit the four labs founded in 2023) to detect distillation attempts. Congress is now pushing new controls on the equipment that makes the chips, extending the target beyond the chips themselves. The broader pattern: export controls keep tightening and the gap between the US model lead and Chinese open-weight performance keeps narrowing anyway. Two trends moving in opposite directions.