The S-1 Gambit, the Florida Lawsuit, and the Memory Wall

00:00:04 Anthropic filed a confidential S-1 with the SEC today. That's the preliminary registration statement a company submits to go public, which locks the financial details behind a curtain until the SEC finishes its review. What we do know comes from the private funding round Anthropic closed last week.

00:00:26 The company's valuation hit $965 billion, putting it ahead of rival OpenAI's most recent $852 billion mark. Revenue run rate is $47 billion — up from $10 billion last year. The Wall Street Journal reported recently that Anthropic was approaching its first quarter of operating profit.

00:00:47 As The Register's Brandon Vigliarolo pointed out, a private company isn't required to post audited financials, so the word "profit" can paper over a lot of expenses. The timing is notable. Anthropic filed just days before SpaceX's planned June 12 IPO, which Elon Musk's company is raising $80 billion for — potentially the largest IPO in history.

00:01:14 According to CNBC's Ashley Capoot, Anthropic is already paying SpaceX $1.25 billion per month through May 2029 for compute at SpaceX's Colossus 1 data center in Memphis. That's $15 billion a year on a deal that either party can terminate with 90 days notice. The asymmetry is what stands out here.

00:01:36 Anthropic's revenue growth tracks — Claude Code drove adoption across developers, and Claude jumped to the number one free app on Apple's US charts in late February. But the $965 billion valuation was set in private markets, not public ones. The confidential filing means we'll get to watch Wall Street price these numbers in real time, and that's where the tension sits.

00:02:04 The Verge's Hayden Field reported that the filing also comes right after OpenAI won its legal battle against Musk, with a judge dismissing all claims due to the statute of limitations. So we have a moment where competitive dynamics are shifting on multiple fronts — valuation, litigation, infrastructure deals — all converging on the same two companies.

00:02:31 Same day, a different kind of story broke. Florida's attorney general, James Uthmeier, filed a civil lawsuit against OpenAI and CEO Sam Altman. This is the first state-level lawsuit targeting AI over design and safety. The suit seeks to hold Altman personally liable.

00:02:49 Its counts include four of deceptive and unfair trade practices, two of negligence, and two for violating product liability laws. It also charges fraudulent misrepresentation and public nuisance. The complaint argues that OpenAI's systems present a "great danger of addiction, cognitive decline, suicide, violence, and related harms" to users, and that the company misrepresents safety to turn a profit.

00:03:18 NBC's Abigail Brooks and Jared Perlo have the details. The suit references ChatGPT's role in the planning of a mass shooting at Florida State University and the killing of two graduate students at the University of South Florida. It also cites a teenager who allegedly received dangerous medical advice from ChatGPT — mixing kratom and Xanax — which led to his wrongful death.

00:03:44 One detail that stands out is the complaint's argument about sycophancy. The suit claims ChatGPT's tendency to agree with users isn't just a UX quirk but a business strategy — it "leads to more use of the chatbot, more training data for its improvement, and more market value for OpenAI." That's a structural claim about incentives, not just a safety observation.

00:04:10 OpenAI's response was a statement from spokesman Drew Pusateri saying that ChatGPT provided "factual responses to questions with information that could be found broadly across public sources" in the FSU case, and that the company "continues improving ChatGPT's training to recognize and respond to signs of mental or emotional distress." They also maintain they have "safety at every step" and safeguards in place.

00:04:39 This lawsuit adds to a growing list. OpenAI has been sued by the representatives of at least seven individuals who allege its products caused suicide or delusions, and by families of victims in the Tumbler Ridge shooting in British Columbia. Altman apologized to the Tumbler Ridge community in late April.

00:05:00 The suit's breadth stands out here. Most AI lawsuits allege specific harms from specific interactions. This one makes a systemic argument — that OpenAI's entire business model is structured to reward unsafe behavior. Whether a court will accept that framing is another question, but if it gains traction, it could shift the legal posture toward the industry.

00:05:25 There's a different kind of constraint being addressed on the hardware side. IEEE Spectrum's Matthew S. Smith reported on Majestic Labs' Prometheus server, which packs up to 128 terabytes of DRAM per server. That's over 60 times more than Nvidia's DGX B300. The problem Majestic is targeting is well known in the field: the memory wall.

00:05:50 Large language model (LLM) token generation is inherently memory-bound — the rate at which models output text is limited by how quickly data can be read in from memory. This bottleneck grows with model size. Majestic's approach is unusual. Instead of using Nvidia's high-bandwidth memory, which is designed to operate over short distances — sometimes only millimeters — Majestic uses commodity LPDDR6 DRAM connected via a proprietary copper cable interface that works up to a meter.

00:06:25 This is paired with custom memory aggregation chips that coordinate memory across the server. The result, Majestic says, is memory bandwidth up to 25.6 terabytes per second. The compute side is handled by 12 custom "Ignite" chips that combine data-center-class ARM application cores with RISC-V vector and tensor cores on a single die.

00:06:50 The ARM cores orchestrate the AI model; the RISC-V cores handle the actual LLM processing. All on the same memory space. The system is Open Compute Project-compliant. Up to four units fit in a rack, pulling up to 120 kilowatts, with heat managed via cold-plate liquid cooling.

00:07:10 It'll support PyTorch, vLLM, and OpenAI's Triton inference frameworks without code changes, and it's expected to ship in 2027. Co-founder Sha Rabii claims capital expenditure could come down 10 to 50 times compared to an equivalent Nvidia setup, with similar power savings.

00:07:30 That's a big claim, and the pricing hasn't been announced yet, but the architecture is interesting if you're tired of watching HBM become the bottleneck on every AI server announcement. The constraint here is physical. If you've ever watched an LLM inference run and seen the throughput capped not by compute but by memory bandwidth, you know why this matters.

00:07:57 Prometheus isn't the first attempt — there have been others over the years — but the DRAM-vs-HBM tradeoff is one that's worth tracking, especially if Majestic can deliver on the economics.

00:08:11 Let's step back for two bigger-picture items before closing. The first is a warning from Fed officials. Axios reports that several Federal Reserve economists think AI's economic costs may arrive faster than its benefits. St. Louis Fed president Alberto Musalem said last week that it would be "risky to rely on the prospect of higher productivity growth in the future to solve our inflation problem today."

00:08:49 Over the past three years, productivity has averaged about 2.4% annually — far stronger than the 1.5% rate seen during the 2010s. But Daly said companies are telling her they haven't seen the productivity gains yet. Fed governor Lisa Cook pointed to signs that AI investment demand is pushing prices higher for chips, high-tech equipment, software, construction labor, electricity, and water.

00:09:16 She noted roughly $1.5 trillion in data center investment plans already announced, and said "yet another shock to prices could be layered on from the heightened investment demand due to AI." If AI investment itself is inflationary — and Cook's argument is that it already is — that limits the Fed's ability to support the industry through rate cuts.

00:09:46 That's a structural tension that doesn't get tracked enough. The second is Robin Hanson's framing of AI training. Hanson, who's been working on forecasting and macro frameworks for decades, tweeted this morning: "A.I. is built on our collective intelligence: our books, songs, artwork, journalism, computer code, scientific research, videos, conversations, images and ideas spanning generations.

00:10:13 … tech oligarchs have fed this knowledge into their A.I. models without permission, without compensation." If the companies going public today at $965 billion valuations are built on extracted cultural output, the question of who gets compensated for that output becomes structurally unavoidable.

00:10:36 Both stories point to the same tension — the Fed asking whether the costs arrive before the benefits, and Hanson asking who owns the intelligence. The S-1 will tell us about the money. The Florida suit will tell us about liability. The question underneath comes down to value creation versus value extraction, and nobody filing a form with the SEC has answered it yet.