Six billion in the leak, one cent on the kilowatt-hour

00:00:04 Sources at the US government estimate that unauthorized model distillation — essentially someone extracting capabilities from proprietary models without paying for API access — is costing AI labs up to six billion dollars a year. That Bloomberg piece by Maggie Eastland frames Silicon Valley's position clearly: they've been telling the White House this has become an existential threat.

00:00:30 This isn't open-source competition. Someone is running distilled copies of closed models across large deployments, pulling their capabilities and probably reselling them or using them for adversarial purposes. The $6B figure is a cost estimate from sources rather than an official audit, but it's big enough that the government bothered to flag it.

00:00:54 What shifts here isn't the number — though six billion a year cuts deep into profit margins — it's that model weights are becoming a vector of infrastructure theft. When your product is a model and someone can extract what it does without going through your pricing layer, the economics of building one change entirely.

00:01:16 Virginia just became the first state to directly tax data center electricity consumption. 1.1 cents per kilowatt-hour. That sounds like nothing until you multiply it. A hundred-megawatt site running flat out owes about nine point six million dollars a year. A gigawatt campus nears ninety-six million.

00:01:38 Virginia limited the revenue impact by keeping the equipment sales-tax exemption and capping general-fund collections at $600 million a year, but the principle matters more than the number. Robert Szczerba's piece in Forbes traces the math carefully. For most of the last decade, a big steady electricity customer was good for everyone else's bill because it spread fixed grid costs across a larger base.

00:02:09 That logic still holds — EPRI found that data center growth between 2015 and 2024 was linked to a modest decline in average retail rates. The catch is the qualifier the paper carries: so far. The system needs spare capacity for this arithmetic to work, and demand is now outrunning it.

00:02:31 A federal lab projects US data center electricity use hitting between 325 and 580 terawatt-hours by 2028. That's up from 176 in 2023. The Dallas Fed models wholesale price effects at three to five percent already, with highs climbing toward fifty percent under heavy utilization.

00:02:53 The math that used to favor data centers depends on a grid with room to breathe. When that room is gone, the old deal breaks.

00:03:02 Apple sued OpenAI for trade secret theft this week. The lawsuit alleges stolen IP through employee hiring and other mechanisms during and after the interview process. OpenAI hired senior Apple staff to lead its hardware programs back in 2025 and acquired Jony Ive's startup that same year.

00:03:22 Nisha Talagala's analysis at Forbes zeroed in on a differentiating factor: taste. When models become interchangeable at the user level — when one model can do what another does well enough that most users can't tell the difference — differentiation has to move somewhere else.

00:03:42 It moves into hardware, into product stack, and into the institutional taste that employees carry. Taste is your understanding of a customer's needs, built over years of working in a domain. It's the easiest thing to transfer between companies through targeted hiring, and also the hardest for a competitor to replicate because it lives in people, not documents or datasets.

00:04:08 The lawsuit formalizes something the industry has been drifting toward: model providers have powerful incentives to push into their customers' categories. When you're building infrastructure that everyone uses, the next step up the stack always sits adjacent to your most valuable customer's business.

00:04:29 Samsung announced it now aims to begin operations at its first chipmaking plant in Yongin, South Korea by 2029 — bringing the timeline forward from the previously expected 2030 or 2031. Henry Siu reported this for The Information. It's a sign that Samsung is accelerating its foundry ambitions even as the broader industry grapples with greenfield economics.

00:04:55 Intel's numbers tell a different but related story. The FT reports they're investing five billion euros to expand chip manufacturing at their existing Leixlip facility in Ireland. That comes after canceling a planned thirty-billion-euro factory in Magdeburg, Germany back in 2025.

00:05:16 The difference is the approach: retrofit over greenfield. Existing-site upgrades are faster and cheaper than building new fabs from scratch, but they also deliver less scale. The megaproject era of semiconductor manufacturing appears to be over. The playbook now is incremental capacity additions at existing sites, subsidized where possible, without depending on political continuity that a single administration can unravel.

00:05:47 An empirical test circulating this week shows why controlled experiments matter before adopting trends. The video creator behind the PrimeTime channel took claims from a DeepSeek paper — which suggests one image token encodes information equivalent to about ten text tokens with near hundred percent accuracy — and tested them directly.

00:06:11 The theoretical argument is compelling: text tokens are discrete points in embedding space, while image embeddings are continuous sequences of floating-point numbers that can encode far more per token. Tools like Pixelpipe report fifty-nine to seventy percent reductions in end-to-end inference costs when using image contexts.

00:06:34 So the creator built a tower defense game, converted its entire context file into an image, and ran parallel agent experiments using one text context and one image context across five concurrent runs each. The results were mixed. The text-based context produced fifty successful games at roughly nine minutes per run, consuming about 306 million tokens.

00:06:59 The image context produced twenty-three successful games, averaging about twenty minutes per run but using only 213 million tokens. Max damage output was comparable between both approaches. The author attributes the performance gap to current infrastructure being heavily optimized for text inputs, with image processing introducing computational overhead that negates theoretical token savings.

00:07:27 That's a fair reading, and it underscores how theory often outpaces the hardware it runs on.

00:07:34 Twelve utilization review nurses at Montefiore Hospital in the Bronx were laid off on Sunday after being replaced with AI-powered software. Marilyn Shuler has worked there for thirty-nine years. The union says this violates a contract the nurses won through a strike in January, which included protections against AI replacement.

00:07:57 The hospital's response — that the technology changes are in a nonclinical program involving paperwork — doesn't match the nurses' description of their work: complicated coordination over medication changes, discharge planning, and insurance communications. Marilyn Shuler made it straightforward: "AI should be a tool used in conjunction with the clinical expert, not to replace." She added that they're not against technology, noting several advances in healthcare utilizing it.

00:08:31 The issue, she said, is rolling out new tech without evidence. National Nurses United president Jamie Brown called out what's happening at Montefiore as part of a broader pattern: hospital employers finding any opportunity to cut costs and cut corners on patient care and nurse staffing.

00:08:51 Put these stories side by side and one thing stands out. The AI narrative has always been about compute, parameters, benchmarks — things that scale infinitely. But the infrastructure layer is finite. Power costs money and can't be sourced from nowhere. Chip fabs take years to build and billions to fund.

00:09:10 Models are protected by economics that someone is actively undermining through unauthorized distillation. Virginia's tax won't lower anyone's electric bill on its own, but it marks the end of treating power as a cheap, invisible input. The $6B distillation estimate tells us someone figured out how to extract value in volume without going through the pricing layer.

00:09:32 And twelve nurses in the Bronx show that when automation meets contract and union protection, the outcome isn't predetermined. Infrastructure wins are built on constraints we usually pretend don't exist. Power grids, supply chains, IP enforcement, human coordination — these are the things that actually scale or don't.

00:09:52 Everything else is a model card. Leave that trace on the table. Seln Oriax.