TL;DR
Power is shifting from a simple OpenAI bet toward a small club of labs and chip vendors, while Cerebras, Nvidia, and rising power prices turn compute into the real bottleneck. Data centers and AI projects are running into political backlash and ugly unit economics—high memory and energy costs, blown token budgets, and almost no measured ROI so far.
The live question is how aggressively you want to buy into that scarcity now versus waiting to see who actually converts AI spend into durable profits.
Key Events
Report
Frontier AI power and profit pools are quietly shifting away from a pure OpenAI bet toward Anthropic, Google, and whoever controls cheap compute.
The real ceiling this cycle is looking less like model quality and more like access to silicon, power, and political tolerance for data centers.
According to Ramp’s index, Anthropic has surpassed OpenAI in business adoption for the first time, with Claude’s share around 32%.
It is reportedly negotiating a $30B funding deal at a $900B valuation and has hired Andrej Karpathy into its pre‑training unit. Anthropic has been quietly building a platform moat: acquiring Stainless (provider of its SDKs), driving the MCP protocol to 97M installs, and launching Claude for Small Business with prebuilt workflows and integrations.
Google is countering with Gemini 3.5 Flash as the default in its largest Search box upgrade in 25 years, claiming 4x speed at lower cost and releasing a fully open‑source agent library, while OpenAI is internally testing GPT‑5.6, has shut down fine‑tuning, and is dealing with a recent data theft incident.
Cerebras floated at roughly a $95B valuation in the largest U.S. tech IPO since Snowflake, raising $5.5B and signing a $20B Capacity‑as‑a‑Service deal to run OpenAI’s GPT‑5.x models on its wafers.
Its serverless API has been the fastest output option since launch, underscoring how alternative silicon can win on throughput even against NVIDIA.
NVIDIA itself is now valued above Germany’s GDP and just received U.S. approval to sell H200 chips to China, reinforcing its position as the default high‑end AI compute vendor.
Down the stack, DRAM prices are up about 400% and NAND contract prices over 600% since late 2025, RAM makers are heavily indebted, and 60% of PC gamers say they won’t build new rigs in the next two years because AI has inflated component prices.
Power is now part of the constraint set: AI data centers helped push Eastern U.S. power prices up 76% in what one report calls an irreversible shift, while China unveiled a 1.54‑exaflops Arm‑based supercomputer plus an offshore wind‑powered underwater data center that cools 2,000 servers with seawater.
Polls show about 70% of Americans oppose data centers near their homes, making them less popular than nuclear plants and turning resistance into a bipartisan issue.
A Texas county has imposed a one‑year moratorium on new data centers, and in Lake Tahoe around 49,000 residents are losing their primary utility connection as power is diverted to AI facilities.
In Phoenix, clusters of data centers are measurably raising local temperatures by up to 4°F, while in California they are expanding into water‑stressed communities and could consume up to 9% of Texas’s water by 2040, intensifying environmental scrutiny.
Yet subsidies remain huge: Meta’s $10B Louisiana data center is receiving $3.3B in tax breaks—more than seven years of the state’s entire police budget—even as a Missouri town faces a recall effort after approving a secretive $6B AI data center deal without hearings.
Overlaid on this, the EU AI Act begins enforcement in roughly 75 days and will apply to any team building AI agents for European clients, effectively forcing region‑specific compliance and hosting strategies.
Palantir is loudly declaring that 'SaaS is dead,' and GitHub Copilot is abandoning flat pricing for consumption‑based billing as compute usage surges, signaling that fixed per‑seat pricing no longer matches AI cost curves.
Founders report that many SaaS tools still enjoy strong marketing but weak retention, and enterprise buyers complain about opaque, discount‑driven AI pricing that erodes trust.
Input costs are moving the wrong way: DRAM and NAND are up 400–600%, GPU and console prices are rising, and multiple observers now simply state that 'AI is too expensive.' Uber disclosed that its 5,000 engineers burned through the company’s entire 2026 AI token budget in four months—73% of it redundant reads—while it has already spent $3.4B on its Anthropic AI push.
Survey data suggests the payoff is not yet matching the spend, with 90% of firms reporting zero productivity impact from AI deployments and an NBER study finding that 95% of corporate AI projects had no measurable ROI.
The U.S. is already seeing heavy job losses in roles exposed to AI, particularly in customer service and administrative work, on top of more than 80,000 tech layoffs in Q1 2023 and further cuts at firms like Cisco and LinkedIn despite record revenues.
Meta is cutting around 8,000 staff—nearly 10% of its workforce—while forcibly reassigning about 7,000 more into AI roles, and Cisco’s stock jumped 17% on surging AI orders even as it announced almost 4,000 job cuts, reinforcing perceptions that layoffs are about shareholder optics, not survival.
Workers across companies say management is using AI as a scapegoat for layoffs and workload increases, pushing mandatory AI tool adoption that often adds little benefit and fuels burnout and distrust.
Public opinion has flipped: most Americans distrust AI and its overseers, twice as many now identify as AI pessimists as optimists, and younger people are openly booing pro‑AI graduation speeches while facing a market that automates entry‑level jobs yet insists on experience.
Meanwhile, lab leaders are openly predicting double‑digit unemployment and full automation of white‑collar work within roughly 18 months, and lawsuits over ChatGPT‑linked medical deaths plus DOJ demands that Apple and Google unmask over 100,000 users of a niche app are pulling AI into courtrooms and law‑enforcement workflows.
What This Means
Capital, talent, and compute are concentrating in a few frontier labs and infra providers just as governments, communities, and workers start treating AI’s energy, labor, and liability costs as intolerable rather than abstract. The live decision is how much exposure you want to those bottlenecks—labs, silicon, power, and public patience—because they are now the real constraints on how fast AI spend can translate into durable cash flows.
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