CxO · Weekly · April 27, 2026

the hard cap on compute and power

Data centers already consume roughly 7% of U.S. power.[3] Half of the AI facilities planned for 2026 are delayed or cancelled because transformers are scarce and expensive.[4]

Transformer prices have roughly tripled in four years, turning basic grid equipment into a strategic asset for anyone scaling AI.[5] Gas projects for just 11 data‑center campuses could emit more greenhouse gases than entire countries.[6]

One study pegs the hidden environmental and health cost of data centers at $25B a year.[7] Local politics is starting to bite, with Maine vetoing some new centers but predicting a moratorium will pass, and cities like Monterey Park outright banning them.[8]

Chipmakers themselves expect to cover only about 60% of AI memory demand by 2027, so even if GPUs are available, feeding them becomes the next choke point.[9]

capital piling into a tiny AI cartel

Anthropic has effectively become a third hyperscaler, taking a $5B investment from Amazon tied to a long‑term cloud deal.[10] That agreement commits Anthropic to roughly $100B of cloud spend over the next decade on Amazon’s Trainium‑based compute.[11]

Google is hedging too, planning to invest up to $40B into Anthropic on top of its own Gemini stack.[12] Secondary markets now price Anthropic at around a trillion‑dollar valuation, above OpenAI, even as it raises prices to manage ongoing compute shortages.[13]

In parallel, Jeff Bezos’ Project Prometheus raised about $10B at a ~$38B valuation and SpaceX secured an option to buy coding‑assistant startup Cursor for roughly $60B, concentrating enormous capital into a handful of AI platforms.[14]

the chinese stack: cheap, big, and good enough

DeepSeek V4 Pro and Flash run on Huawei’s Ascend 950, offering a 1M‑token context window and parameter counts in the trillion range.[15]

The Pro and Flash variants are priced as low as $1.74 and $0.028 per million input tokens, with a temporary 75% API discount undercutting Western frontier models.[16]

Huawei’s domestic chips have already captured about 41% of China’s AI server market, pairing these models with a homegrown hardware base.[17] Chinese systems like Xiaomi’s MiMo‑V2.5‑Pro now match U.S. frontier coding AIs on SWE‑Bench Pro while using 40–60% fewer tokens, and open‑weight Chinese models increasingly beat U.S. ones on cost.[18]

All this lands against U.S. accusations of "industrial‑scale" AI theft and open talk of an AI Cold War, so the cheap Chinese stack is advancing technically while becoming geopolitically contentious.[19]

labor, surveillance, and the AI blame game

Meta is cutting about 10% of its workforce, roughly 8,000 people, while rolling out software that captures mouse movements, keystrokes, and screenshots from employees to train its AI.[20]

Microsoft is offering voluntary buyouts to up to 7% of its U.S. workforce and plans to cut around 8,750 jobs as part of its AI spending pivot.[21]

IBM has already replaced about 43,000 American staff with 135,000 workers in India, and a study projects over 260,000 people in Massachusetts alone could lose jobs to AI systems within five years.[22]

At the same time, UK data show no significant impact of AI on overall employment with AI‑exposed jobs actually growing faster, and thousands of CEOs admit their AI efforts have had little effect on productivity or headcount.[23]

Workers are responding with open pushback, from Samsung’s 30,000‑plus union members demanding roughly $400k bonuses each from a $38B AI memory windfall to employees framing layoffs as profit grabs rather than technological inevitability.[24]

ROI reality check on AI agents and code

Practitioners estimate that maybe 20% of deployed AI agents deliver real ROI, with the rest mostly hype or shallow time‑savers.[25] Discussions highlight that classic predictive models often beat flashy generative agents on return, and that real value tends to come from narrow, recurring workflows rather than grand "autonomous employee" visions.[26]

Google now says about 75% of its new code is AI‑generated, up from 25% in 2024, but users are loudly complaining about deteriorating quality in products like Search and Maps.[27]

A controlled study found that when 1,222 people lost access to AI assistants, their performance fell below a control group, yet benchmarks like CI‑Work show enterprise agents still leak sensitive data in 15.8–50.9% of test cases.[28]

There are genuine wins—Starlink replaced its call center with an agent called Grok that now resolves roughly 70% of calls, and the UAE wants half of government operations running on agentic AI within two years—but the hit rate remains low.[29]