Anthropic's $900B moat, Cerebras pops 68%, and Hashimoto calls 'AI psychosis'
Mon, May 18, 2026 · 10 stories
AI Psychosis is a delusion of accepting and shipping lower-quality code or deprioritizing resilience because AI Agents can fix bugs rapidly and at scale. Mitchell Hashimoto (co-founder of HashiCorp) argues that even though the MTTR (mean-time-to-recovery) is lower the underlying architecture is decaying. You may have full test coverage, but an unrecognizable architecture is a disaster waiting to happen.
Amazon and other large employers are heavily pushing employees to adopt AI-powered tools. Amazon told staff AI usage statistics won't factor into performance evaluations, but it's hard to believe managers won't be checking. Internal tooling that automates repetitive tasks like monitoring deployments during meetings or triaging email before you wake up conveniently lifts the same usage numbers nobody is supposedly tracking. Meta is running a similar play.
Compute is not cheap. VCs are footing the bill for AI compute. This isn't an AI bubble, it's going to be an AI bomb. These frontier models are very power hungry and the subscriptions and API pricing we're seeing are highly subsidized by VCs. Investors eventually want their money back and when they call, the user base is going to have to answer.
Top Stories
Bloomberg Technology / TechCrunch AI
Anthropic is in early talks to raise approximately $30 billion at a $900 billion valuation, with the company simultaneously warning investors against unauthorized secondary platforms offering exposure to its shares. The same week, third-party Ramp spend data showed Anthropic has surpassed OpenAI in number of business customers, while Anthropic publicly expanded its push into the legal industry with vertical AI tooling.
Why this matters:
The Anthropic commercial moat thesis we've been tracking just crystallized in one week. Last week we noted Anthropic was raking in private capital while losing Pentagon contracts to OpenAI, Google, and Microsoft. This week, independent transaction data shows them quietly winning the enterprise developer wallet, and a fresh $30B round prices that lead at $900B. For the AI founders and VCs reading this, the model-picker conversation just shifted from "OpenAI default, Anthropic alt" to "Anthropic for serious enterprise, OpenAI for consumer scale." Expect Anthropic's pricing posture to harden over the next two quarters.
Bloomberg Technology
Cerebras Systems raised $5.55 billion in 2026's largest IPO and shares surged 68 to 89 percent in trading debut, minting a $3.2B fortune for the CEO and billion-dollar exits for early backers Benchmark and Eclipse. The same week, Bloomberg disclosed Microsoft has spent over $100 billion to date on its OpenAI partnership, and OpenAI's CFO publicly told investors the company may need to raise additional capital as its compute crunch deepens. CME separately announced a futures market for AI computing power, financializing the bottleneck.
Why this matters:
Three stories that together name the bottleneck: AI compute is now scarce enough to be priced in public markets, securitized as a futures contract, and forcing the largest tech investment in history to keep growing. Cerebras' 89 percent pop is the cleanest read yet on the public market's appetite for AI infrastructure exposure. Microsoft's $100B+ disclosure recasts the OpenAI deal as comparable to landmark industrial-era bets. For builders, the practical implication is that the era of "compute is cheap, optimize for developer velocity" is over. Expect rate limits to tighten, enterprise contracts to require capacity commitments, and inference pricing to bifurcate between hyperscale customers and everyone else.
X (Mitchell Hashimoto), Ars Technica, The State of Brand
Three unrelated pieces from three sources hit Hacker News within four days and all centered on the same theme: enterprise AI adoption is breaking down at the operator level. HashiCorp co-founder Mitchell Hashimoto posted on X that he "strongly believes there are entire companies now under AI psychosis" (594 HN upvotes, 268 comments, the week's highest community engagement). Separately, Ars Technica reported on internal Amazon practices of "tokenmaxxing," where employees inflate AI tool usage metrics to satisfy mandate pressure rather than to do useful work. And an enterprise-cost essay in The State of Brand arguing "every AI subscription is a ticking time bomb" pulled 280 HN points the same week.
Why this matters:
Three independent voices, a respected founder, frontline engineers at the world's largest employer, and an enterprise-cost analyst, surfaced the same complaint within a week. None of these stories would matter alone. Together they're the strongest counter-signal to this quarter's capital euphoria, and they came from operators, not pundits. For VCs, this is the early warning that the next 18 months will separate companies actually deriving AI value from companies generating impressive-looking usage dashboards. For founders selling AI to enterprise, expect procurement to start demanding usage-tied pricing and outcome guarantees by Q4.
Quick Hits
Bloomberg Technology
Defense AI capital flood continues; connects to last week's Pentagon contract awards that excluded Anthropic.
Bloomberg Technology
Musk's lab enters the coding-agent race already crowded with Claude Code, Codex, and Cursor.
Interconnects
Nathan Lambert's weekly synthesis catalogs the open-weights cycle visibly accelerating, with CAISI's V4 safety assessment as the policy subplot.
Bloomberg Technology
Workflow-automation infra is being priced as picks-and-shovels for the agent era.
Bloomberg Technology
Anthropic's restricted-access program is now visible as a market gap competitors are filling.
TechCrunch AI
Free dictation on the keyboard from Google lands directly on a $2B funding round for a dictation startup.
kabir.au via HN
A technical post arguing frontier models solve cybersecurity capture-the-flag challenges at superhuman levels, making human-only competition obsolete.
