OpenAI confidentially filed its S-1 eight days after Anthropic, at an $852 billion valuation and roughly $2 billion a month in revenue. The same day, Altman and Pachocki published a mission statement that reads exactly like the prospectus narrative they will need. Underneath the IPO noise, three independent studies converged on the same uncomfortable finding: the model layer is no longer the constraint, the workflow around it is, and the trillion-dollar lab valuations depend on a moat that may have already moved.
The IPO Queue Gets Specific: $852B, A Mission Statement, And A 2028 Holdout
OpenAI joined Anthropic in confidential filing, Altman and Pachocki published the prospectus narrative as a mission essay, OpenAI opened an external economic research program, and Perplexity locked in 2028.
- OpenAI files confidential S-1, eight days after Anthropic — $852 billion valuation at filing. Goldman Sachs and Morgan Stanley underwriting. Around $2 billion a month in revenue, $13.1 billion last year, still unprofitable. Two trillion-dollar AI lab IPOs queued in eight days. The public-market AI era is officially active.
- OpenAI’s mission statement reads like an S-1 — Altman and Pachocki published three goals on the day before the filing announcement: an automated AI researcher by March 2028, accelerating the economy with shared gains, and a personal AGI for everyone. Compare it to Anthropic’s pause-button essay last week. Two trillion-dollar labs, opposite policy postures, same investor audience.
- OpenAI opens a research program on AI’s economic effects — External academics, $250 million committed by the OpenAI Foundation, RFP closes July 5. Third piece of OpenAI’s IPO-prep policy push, after Public Wealth Fund discussions and the Built to Benefit Everyone plan.
- Perplexity locks in 2028 IPO, skipping the first AI listing wave — CEO Aravind Srinivas told CNBC the 2028 timeline holds regardless of how Anthropic and OpenAI price. The first major AI-native consumer company to opt out of the 2026-2027 window. Confidence or patience depends on what the S-1s eventually disclose.
The Bottleneck Quietly Leaves The Model Layer
Three independent studies and one essay all argued the same thing this week. The model is no longer the constraint. The workflow around it is.
- The workflow is the moat, and that breaks the lab-IPO thesis — K-Dense’s argument: Claude Opus, GPT-5.5, Gemini 3.5 are all production-acceptable. The differentiation has moved to prompt construction, retrieval, tool selection, evaluation. If workflow is the moat, the trillion-dollar lab valuations need a different premise.
- The 8% reality: what AI coding tools actually deliver — DX’s longitudinal study finds median PR throughput gains of 8% across the broad adoption population, with upper-range 10-15% confined to orgs that redesigned reviews, planning, testing, and coordination. The empirical counterweight to Anthropic’s 8x velocity claim from last week.
- Perplexity’s agent study puts 87% and 94% on the procurement deck — Perplexity Research reports 87% time reduction and 94% cost reduction using the Computer agent on research-and-synthesise tasks. The numbers are real and vendor-published. They are also task-category dependent. Pair them with the DX 8% finding and the honest read emerges.
- The workflow that fixes itself is the deepest moat — Avi Chawla argues agent observability has been stuck at trace level for a year. The next architectural shift is harnesses that diagnose their own failures, propose patches, and verify fixes without human intervention. The pieces are now lining up: sandboxes, pattern surfacing, agent identity, self-repair.
Xiaomi Hits A Trillion Parameters At 1,000 Tokens A Second. xAI Becomes A Landlord.
A consumer-electronics giant shipped a frontier-speed model on commodity hardware while the most strategic compute story of the week was xAI quietly becoming infrastructure with an AI brand attached.
- Xiaomi hits 1,000 tokens/sec on a trillion-param model using FP4 and DFlash — Xiaomi and inference partner TileRT shipped MiMo-V2.5-Pro-UltraSpeed: 1-trillion-parameter MoE, FP4 quantization on expert layers, DFlash speculative decoding, 1,000 tokens per second on a standard 8-GPU node. Three times the price for roughly 10x throughput. Latency-sensitive agent workflows have a new option.
- xAI is now a datacenter landlord with an AI brand attached — Martin Alderson’s analysis: 18-month payback on xAI’s full capex from the Anthropic and Google capacity-rental deals. 90-day cancellation clauses. Hundreds of MW of capacity still available. xAI’s frontier-lab positioning is increasingly window dressing on a profitable infrastructure business.
- Starlink operators are taking over xAI’s engineering ahead of the SpaceX IPO — Bloomberg reports a Starlink customer-support engineer is replacing a college-age frontier-AI engineer to lead Grok training. The Starlink-to-xAI personnel pipeline is a corporate restructuring. Musk is presenting xAI to public markets through the SpaceX infrastructure lens.
Today’s Quick Hits
- SchemaFlow turns DB change requests into a six-layer AI workflow — OpenAI partner SchemaFlow published the reference architecture for AI-assisted database changes: structured request parsing, impact analysis, SQL generation, guardrails, artifact creation, evals. The agent stack is gaining specialised verticals.
- Cognition’s FrontierCode asks if a model’s PR would actually merge — The first code benchmark to measure acceptability rather than correctness. Built by open-source maintainers, judging by what gets merged in production. Strategic positioning that favours Devin and also a useful industry artefact.