CxO · Weekly · June 10, 2026

the gpu megadeals and frontier ipo layer

Google agreed to pay SpaceX/xAI about $920M per month for compute through June 2029, buying access to roughly 110,000 NVIDIA GPUs.[2]

Anthropic is reportedly paying SpaceX around $1.25B per month for GPU capacity, while separately inking a roughly $15B‑per‑year compute deal with Google.[3]

At the same time, SpaceX is targeting a $1.75T IPO that would require ~60x growth in a decade and a 100‑fold increase in AI revenue by 2030, projections that analysts and Morningstar already describe as overvalued and risky for pension funds.[4]

Commenters tracking these deals argue that key GPU contracts may be margin‑negative with termination clauses, making SpaceX and xAI look as much like leveraged data‑center REITs as high‑margin AI platforms.[5]

frontier labs: indispensable vendors, dubious securities

Anthropic and OpenAI have both filed for US IPOs after being rejected from early S&P 500 inclusion, with valuations around $965B and $852B respectively despite heavy operating losses and burn rates estimated at over $1000 spent per $100 of revenue.[6]

Anthropic reports that more than 80% of the code merged into its own codebase is now written by Claude, and its new Claude Fable 5 model scores at senior‑engineer level on benchmarks while speeding up training code by up to 52x.[7]

Skeptical investors and developers compare these sky‑high valuations unfavorably with hardware firms like ASML and see Anthropic’s high‑profile calls for a global AI pause as timed marketing to support its IPO narrative rather than purely safety‑driven.[8]

There is a widely voiced concern that the upcoming AI IPOs are structured to offload risk onto retail and pension investors, with growth expectations that many commenters describe as closer to hype than to credible cash‑flow forecasts.[9]

regime divergence: us preemption vs eu sovereignty and unbundling

US House lawmakers have floated a bill that would prohibit state‑level AI regulations, effectively centralizing oversight in Washington and signaling a relatively light‑touch, single‑regime environment for AI businesses.[10]

In contrast, the EU is formalizing a Cloud Sovereignty Framework that scores ‘digital sovereignty’ and is investing in an Open Source Strategy to reduce dependence on US tech stacks.[11]

EU regulators denied Apple an exemption from tech rules, leading Apple to withhold its revamped Siri AI from the EU and delay parts of iOS 27 and iPadOS 27, which has EU users complaining they receive inferior products.[12]

The EU also ordered Meta to allow rival AI chatbots to operate on WhatsApp for free, a direct forced‑interoperability move against a dominant messaging platform.[13]

infra limits: water, grids, bans, and nuclear workarounds

US data centers consumed about 264B gallons of water under drought conditions affecting nearly 63% of the country, and roughly two‑thirds of 809 planned new US AI data centers are sited in drought‑hit zones.[14]

Spending on data centers in the US now exceeds federal transportation spending, while public opposition—especially in working‑class neighborhoods—is driving moratoria and tax‑break suspensions.[15]

New York has imposed a one‑year ban on large data centers over 20MW, Seattle is moving toward banning new data centers entirely, and Amazon employees in Seattle are demanding a one‑year moratorium on local data‑center projects.[16]

Forecasts suggest that by 2030 AI’s water use could match that of 1.3B people and its power draw could triple the usage of 650M people, pushing operators toward new energy architectures.[17]

China has already deployed a nuclear reactor designed to power data centers for decades on a single fuel load and brought a 24MW underwater, wind‑powered data center online off Shanghai, while startups like Antares Nuclear are achieving criticality on new reactor designs aimed at similar markets.[18]

agents now dominate traffic, but security and trust are cracking

Cloudflare reports that bots and AI agents have, for the first time, surpassed human traffic online, with agentic traffic growing faster than expected.[19]

AI agents are generating more web traffic than humans overall, and usage patterns have shifted toward long‑running agents that drive up token consumption and costs.[20]

In production, a LangChain‑based customer‑support agent failed in about 30% of user sessions, McDonald’s AI drive‑thru faced customer backlash, and mobile agent apps are proliferating without matching user engagement.[21]

Security incidents are already significant: Meta confirmed that abuse of its AI chatbot led to roughly 20,000 Instagram accounts being hacked, researchers at the University of Toronto demonstrated an AI worm that could target any online device, and AI‑generated scams cost Americans nearly $900M last year.[22]

Regulators are responding piecemeal, with New York requiring AI disclosures in news content, English and Welsh police halting AI use in court statements, and UK banks identifying AI cyber‑risk as a primary systemic threat.[23]