TL;DR
Money and attention shifted toward alternatives to Nvidia and OpenAI, from Cerebras’ blockbuster IPO to Anthropic taking the lead in enterprise AI. Governments tightened their grip on data and infrastructure, squeezing U.S. hyperscalers while subsidizing massive new AI builds and tolerating ugly labor fights.
The game is less about picking a single winner now and more about navigating a fragmented, politicized stack.
Key Events
Report
Capital is rotating from pure-model bets into whoever controls scarce, differentiated compute and regulated data. At the same time, the model leaderboard is reshuffling, and the business models on top of it look a lot less bulletproof than the hype.
Cerebras came public around a $95B valuation, the largest U.S. tech IPO since Snowflake. It raised $5.5B, and demand for the deal far outstripped available shares.
Its chips are already serving GPT5.4 and GPT5.5‑class models, with GPT5.5 estimated at around 2T parameters, and early users say inference is 'exceptionally fast' even though new sign‑ups are closed.
AMD’s EPYC has quietly climbed to a record 46.2% server revenue share, and AMD is opening dev clusters that used to be NVIDIA‑only territory while cutting mid‑range GPU prices by roughly high‑single digits.
At the edge, NVIDIA is pushing a $249 desktop AI box that can run LLMs locally, while Intel’s Optane rigs are being shown running 1T‑parameter models at over 4 tokens per second.
MIT’s ‘implosion carving’ optical work adds a speculative, longer‑dated path to non‑GPU acceleration.
Anthropic has, for the first time, overtaken OpenAI in business adoption and pushed ChatGPT to second place in key gen‑AI metrics, while reporting an annualized revenue run‑rate north of $30B in early April 2026.
It is leaning into verticals with launches like Claude for Legal and Claude for Small Business, plus a $200M partnership with the Gates Foundation to push AI into global health.
Under the hood, there is growing backlash against token‑based billing and price hikes, with small and mid‑sized companies calling Claude and peers economically unviable and questioning whether marketing is outpacing real capability.
One AWS user reported a $30k surprise bill after a Claude model ran without guardrails on Bedrock, and Uber’s CTO is publicly struggling to justify a $3.4B Anthropic‑linked AI budget.
On the other side, OpenAI is winding down its fine‑tuning API while standing up a new $4B corporate AI unit, rolling out a Plaid‑backed personal finance assistant that talks to users’ bank accounts, and cutting a deal with Malta to give every citizen ChatGPT Plus for a year if they pass an AI literacy course—and each move is drawing loud privacy and concentration‑risk criticism.
European policymakers are moving to block Microsoft, Amazon, and Google from handling sensitive government health, financial, and legal data, effectively forcing a shift toward sovereign or domestic clouds for that spend.
The UK has already ripped out Palantir’s refugee‑management stack in favor of an in‑house system it says will save millions, while still granting Palantir unlimited access to NHS patient data for other programs.
German intelligence has chosen a French AI firm over Palantir, and several NHS regions are opting out of Palantir in favor of internal or local systems, showing this isn’t an isolated protest.
At the same time, the European Court of Justice has ruled that Meta must compensate Italian publishers for use of their content, further formalizing local value‑sharing demands.
OpenAI’s country‑level deal with Malta—population‑scale ChatGPT Plus in exchange for literacy training and data access terms—illustrates how some governments will still trade sovereignty for capability if the bundle is attractive enough.
Major cloud and AI infra providers are reporting rising profits but weakening cash flow after capex, suggesting that the current build‑out is burning more cash than headline earnings imply.
Oracle is a clear example: profits are up even as post‑capex cash flow falls, while it touts closed‑loop cooling to defend its AI data‑center water use and sheds staff in a broader wave that has cut nearly 100,000 high‑income tech jobs.
Microsoft’s flagship Kenya AI data‑center project is stalled over power and infrastructure disputes, highlighting real grid and siting limits even when capital is available.
Meta is spending $10B on a Louisiana data center and securing $3.3B in tax breaks, while planning to lay off about 8,000 employees despite posting $26B in Q1 net income and reports of a grim internal culture.
Upstream, Samsung has put chip fabs into 'emergency management mode' and is winding down production ahead of a threatened 18‑day strike over profit‑sharing that could cost an estimated $2B per day during South Korea’s AI chip boom.
What This Means
Capital is no longer paying for a single center of gravity in AI; it’s fragmenting across wafers, CPUs, sovereign clouds, and country‑level deals. The cost of being pinned to any one stack, region, or labor pool is quietly compounding.
On Watch
Interesting
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
Sources
Key Events
On Watch
Interesting