June 22, 2026
Six grid operators just got 60 days to answer for AI's power bill
The megawatt bottleneck in one issue: Baseten raised $1.5 billion at a $13 billion valuation to run AI inference, Anthropic's $1.5 billion author settlement prices training data near $3,000 per book, FERC gave six US grid operators 60 days to fix data center interconnection, and General Fusion squeezed plasma to 8.4 million degrees, all pointing to power, not model quality, as the real limit on what you can build.
This was the week AI stopped being a software story and became a power story. Five events, one theme I will call the megawatt bottleneck: the limit on what you can build now is not model quality. It is whether you can get the electricity to run it. Here is the read across all five pillars.
TECH: AI in production
Google put a new image model straight into products people already pay for.
Google's Nano Banana Pro, its Gemini 3 Pro Image model, is built to render legible text inside generated images, which older models botched. It is not stuck in a demo. Google wired it into the Gemini app, Google Ads, and Google Slides.
The money play: If your product has any spot for a generated graphic, ad, or diagram, test this model against what you use now before your next sprint. Text-in-image was the failure point that forced manual design work. That gap just narrowed.
Sources: Google | Google DeepMind
BIZ: build, fund, scale
Baseten raised $1.5 billion to run other companies' AI, at a $13 billion valuation.
The round was reported June 18 and closed near June 22. Baseten sells inference, the compute that runs after a user hits send. Five months earlier it raised $300 million at $5 billion. The bet from investors: inference, not training, is where the money and the margins now sit.
The money play: If you ship AI features, inference is your real cost line, not the subscription to a model maker. Route each request to the cheapest model that clears your quality bar, and keep open-source options in the mix. That is the exact playbook the buyers just funded.
Sources: TechCrunch | Baseten
LAW: rights and fights
The bill for training on other people's work is now a real number: about $3,000 per book.
In Bartz v. Anthropic, the company agreed to pay $1.5 billion to authors and publishers whose books were used, covering roughly 500,000 works. First payments to authors are expected in mid 2026. Separately, on June 11 the Third Circuit heard the first appeals-court argument on whether training AI on copyrighted work is fair use, in Thomson Reuters v. Ross Intelligence. That ruling will set the price for every lab after this one.
The money play: If you fine-tune on scraped data, put a licensing line in your budget now. If you make books, art, or code, check the class lists and file your claim before the deadline passes.
Sources: Authors Guild | LawSites
POL: policy, tax, borders
FERC gave all six US grid operators 60 days to justify or rewrite how data centers plug in.
On June 18, federal regulators issued show-cause orders to PJM, MISO, SPP, CAISO, ISO New England, and NYISO. One named worry: cost shifting, where regular ratepayers end up subsidizing the grid upgrades a data center needs. This is the first hard federal move on the demand the AI build-out is putting on the grid.
The money play: If you are siting compute or a hosting business, treat interconnection as a real risk, not a formality. The rules are being rewritten this quarter.
Sources: Utility Dive | FERC
SCI: breakthroughs to build on
A fusion startup heated plasma to 8.4 million degrees by squeezing it, no lasers.
On June 22, General Fusion said its LM26 machine reached about 8.4 million degrees Celsius, roughly 0.72 keV, by mechanically compressing a plasma. That was a more than threefold jump in electron temperature. It is a step, not a power plant. The next target is 1 keV.
The money play: Do not wait on fusion for cheap power. It will not run your servers this decade. The near-term edge is efficiency and siting near power you can already get.
Sources: General Fusion
Before you go
The pattern this week: the software got cheaper to run and the electricity got harder to get. If you want help picking where AI actually pays off in your business, that is what I do at nomadtechnologist.com. Kobe reviewed this issue by sleeping through it, which is his rating for most AI news.
Not legal, financial, or tax advice.
Get this every weekday.
Free. 5 minutes. Unsubscribe whenever.