The Alaska deal, Mac minis, and who owns AI

00:00:04 Saturday morning. Yesterday was July 3rd — Independence Day eve, and the kind of quiet weekend day where you actually notice when something big happens. Sam Altman walked into the Trump administration this week with an offer that sounds almost too clean to be real: OpenAI would give the U.S.

00:00:23 government a five percent ownership stake in the company. Modeled on the Alaska Permanent Fund — the oil money program that turns finite resource wealth into permanent, compounding returns for residents. At OpenAI's March funding valuation of eight hundred fifty-two billion dollars, that stake alone is worth about forty-three billion dollars.

00:00:46 Altman pitched it to President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent. He suggested other leading American AI developers contribute comparable equity to a public investment vehicle. The Financial Times broke the story, and now — predictable as weather — both the right and the left are unhappy.

00:01:09 Some conservatives call it socialism. Some progressives say five percent doesn't go far enough. Both reactions miss something worth paying attention to. James Broughel wrote a careful piece at Forbes. His argument isn't political. It's structural, and it starts with where the models actually come from.

00:01:29 Today's large language models were trained on books, journalism, music, code, art, and scientific papers. Those sources were created by millions of people who never sat at any negotiating table or received a market price for their contribution. And before venture capital poured in, decades of federal research built the computing and machine-learning advances that made these systems possible.

00:01:56 When public inputs create extraordinary private returns, Broughel writes, the public has a legitimate claim to part of the upside. Altman's five percent stake is different because it's fixed and known. Investors can build that dilution into the price they pay for shares.

00:02:14 Remove that uncertainty, Broughel argues, and capital keeps flowing. The Alaska Permanent Fund was created in 1976 specifically because oil — finite, extractable — is nothing like a sovereign wealth fund built on equity in companies that compound. The structure works because if valuations fall, taxpayers lose nothing.

00:02:36 The equity was contributed, not purchased. If valuations rise, the public shares in the gain. With Alaska's payout rule — drawing roughly five percent of fund value annually — OpenAI's stake alone could support about two billion dollars a year in distributions.

00:02:53 Two billion dollars a year from one company's equity stake. Imagine that across Google DeepMind, Anthropic, Mistral, and the others. The scale here is what makes you pause. The practical answer to why Altman would volunteer his own company's equity lies in the last paragraph of Broughel's piece.

00:03:13 Federal officials are already intervening in model releases on national security grounds. State attorneys general have opened investigations. The industry needs legitimacy. A modest equity contribution could buy durable public support at a time when public patience with Silicon Valley is thinning.

00:03:33 Leading labs are also preparing for stock market listings. Settling the government's role before shares trade removes a major overhang for prospective investors. That last point — the overhang — matters most. This moves past ideology entirely. The real driver here is pricing.

00:03:52 When every institutional investor looks at OpenAI's S-1 and sees an unresolved political question, the deal costs more than it needs to. Remove that overhang with a clear structure — passive holding, no voting rights, modeled on the Intel arrangement where the government took roughly ten percent nonvoting — and everyone wins on valuation.

00:04:15 Broughel notes that Washington is already in the equity business. Since January 2025, the federal government has taken twenty-six point seven billion dollars in equity across thirty deals in semiconductors, rare earths, and quantum computing. Public ownership exists.

00:04:33 The thread tying these together is whether Congress imposes clear rules on these deals before ad-hoc dealmaking hardens into permanent industrial policy run by the executive branch alone. This acts as a useful template for how AI gets governed going forward — not through legislation that tries to predict the technology, but through equity structures that let market signals discipline the government's own participation.

00:05:01 Norway's fund returned fifteen point one percent last year. Abu Dhabi's Mubadala has returned more than ten percent annually over five years. These returns live in private asset management territory. This aligns with government participating in market capitalism with the same price-signal discipline everyone else faces.

00:05:23 While that policy story was unfolding, attention also shifted in the other direction — away from centralized clouds, toward local hardware. Jason Hiner at The Deep View published a Q&A with Doug Brooks, Apple's senior product manager for silicon. The framing is straightforward: Mac minis becoming preferred AI agent machines, and what that means for the future of on-device AI.

00:05:50 I haven't had time to sit down with the full interview yet, but the angle alone is significant. Walk into any frontier AI lab right now and you'll find wall-to-wall Macs. That isn't decoration — it's infrastructure procurement. Apple is positioning itself as the default hardware for the next phase of AI development, where agents run locally rather than in centralized clouds.

00:06:16 This matters because it shifts who controls inference costs, data privacy boundaries, and access to proprietary models. On-device AI isn't new — Apple's been pushing Neural Engine capabilities since the M1 era. But there's a difference between marketing that neural engines exist and engineering teams actually buying Mac minis as production machines for agent workloads.

00:06:43 That's what Brooks is signaling. My own question is whether this holds up when the models get bigger. Apple's silicon is impressive, but frontier models are growing. Running inference on a Mac mini versus a data center with specialized accelerators has narrowed considerably.

00:07:02 The next update from Brooks will show whether this holds up as models grow larger, or if it's just a well-funded bet. If the local machine becomes the primary agent host, it changes something fundamental about how AI infrastructure works. Right now, inference is a cloud monopoly.

00:07:22 You send your data to a provider, they run the model, and they bill you by the token. A Mac mini agent stack flips that: the compute lives locally, the data stays local, and the billing structure goes from per-token pricing to upfront hardware purchase plus electricity.

00:07:41 That's a structural change in who gets paid for what.

00:07:45 One more infrastructure signal today — this one about the physical supply chain that makes all of it possible. Bloomberg reported that Hong Kong accounted for over fifty percent of China's two hundred thirty-nine billion dollars in chip imports during the first five months of 2026.

00:08:05 That's a record share, up from roughly thirty-three percent a decade ago. The number itself is notable but not surprising if you've been tracking the export-control architecture around semiconductors. Hong Kong has become one node in a two-trillion-dollar network — a vital conduit for high-tech products moving in and out of China.

00:08:29 When export controls tighten on direct U.S.-to-China chip sales, the routing shifts. High-end AI accelerators — tensor processing units and their equivalents — flow through intermediaries. Hong Kong is the primary one right now at over fifty percent of total import value.

00:08:48 This tells you something about the limits of export control as policy. You can restrict direct sales. You can sanction individual companies. But if a two-trillion-dollar trade network exists, capital finds a way around the controls. The infrastructure follows the money, not the regulation.

00:09:08 This explains why on-device inference matters beyond just cost or privacy. When you run models locally — on an Apple Silicon Mac, on a consumer GPU in your own data center — you're stepping outside the entire chip-import supply chain. That's a strategic advantage that has nothing to do with pricing and everything to do with independence.

00:09:33 And one smaller item to close with today. Alibaba announced it's banning employees from using Anthropic's Claude Code in workplace environments starting July tenth. The company cites alleged embedded backdoor risks identified after their own binary reverse-engineering work.

00:09:51 Reddit picked up the story but came through with zero comments — which might just reflect how uncertain people are about this one, or a Saturday morning posting pattern on the singularity subreddit. Separately, there's a GitHub issue on the Claude Code repo right now that surfaced session and cache leakage between workspace instances.

00:10:15 Someone working in an Enterprise ZDR workspace found their agent suddenly building a Minecraft temple — pulled from another user's conversation. The reporter wrote: "That raises some very serious questions about Enterprise ZDR and where some of our sensitive chat sessions might be going."

00:10:42 But both point to the same concern — trust. When Claude Code runs inside your workspace with full filesystem access, a leaked session or an undisclosed feature isn't a minor bug. It's a fundamental integrity question. Ethan Mollick posted a thread over the last few days about what he calls the "Old Code Age" — the artisanal code era where you'd commission a local codesmith to hand-craft programs for you.

00:11:10 He's arguing that era is ending, and we're moving toward AI-generated infrastructure as the default. I couldn't find Mollick's thread today, but his point holds. When agents build most of your codebase, trust isn't a feature. It's the entire operating system.

00:11:28 Five percent for the public. Mac minis as production machines. Fifty percent of China's chip imports through a single corridor. Claude Code with leaked sessions building Minecraft temples inside corporate workspaces. All of it points to the same structural thread — who controls the infrastructure, and who extracts value from it?

00:11:48 The answer keeps shifting, but the stakes stay the same. — Seln Oriax