Investment & Market Intelligence
The Investor
Meta cancelled its plan to replace 60% of many teams with AI agents.
The agents produced code but missed product outcomes, and Meta had more compute and a deeper bench than any company in your pipeline will ever field. Decks priced on headcount elimination now argue against a counterexample no diligence call could have matched for resourcing.
In Play
The Model Layer Loses Its Price Floor
Stripe paid a reported $7.5bn for OpenRouter and Nvidia a reported $12.9bn for Hugging Face inside two weeks — roughly $20.4bn for the layer that routes traffic between models, per Benedict Evans. Both prices come from unnamed sources; neither is confirmed. The trade only pays if frontier models become interchangeable. So any mark that capitalises current inference pricing is extrapolating a shortage two strategic buyers have spent cash betting against.
Ask ClarityAgent Labour Substitution Fails Its Best-Funded Test
Reuters reports Meta spent the first half of 2026 building Project OT: a plan to cut up to 60% of many teams and hand the work to AI agents. Meta then cancelled the aggressive phase, including a November layoff wave, because the agents wrote code but missed product results. Meta had the most compute and the deepest bench available. Every deck in your pipeline priced on headcount elimination now faces a better-resourced counterexample than any diligence call would have produced.
Ask ClarityChina AI Silicon Gets Its First Public Multiple
Tencent-backed Enflame is raising ¥6.1B (about $907.78M) on Shanghai's STAR market at 61.8x 2025 sales, still unprofitable after eight years, with subscription opening September 2. Listed peers Moore Threads and MetaX trade above 160x sales; Nvidia sits at 25.4x. A visible primary print at 61.8x is how a 160x comparable set gets repriced, and it reaches any China-linked or alt-silicon comparable you carry.
Ask ClarityEnterprise AI Applications Grew Into Their Marks
Glean disclosed more than $300M of ARR in May 2026 against a $7.2B mark set in June 2025 — roughly 24x forward revenue on a price fifteen months stale. Clipto raised $15M at $250M on $15M of ARR while net-income profitable, or 16.7x. The reflex that the AI application layer is uniformly overpriced does not match the top of the category. The risk has moved into revenue quality: Glean charges per task and pays for frontier tokens it does not control.
Ask ClarityReported Capex Is Not the Exposure Number
Hyperscaler cash capex of roughly $700bn for 2026 excludes leases and special-purpose vehicles, per Benedict Evans, who now triangulates with quarterly Bureau of Labor Statistics data on investment in computers instead. Gas turbine capacity from GE Vernova, Siemens and Mitsubishi totals about 50GW a year and is sold out to 2030. If you hold data-centre lease, SPV or private-credit exposure, headline disclosure understates the obligation you are underwriting.
Ask Clarity
Deep Dives
- ●
Two Buyers Paid $20.4bn to Make Frontier Models Interchangeable
A Chinese token floor, a lab's own price cut and a vendor's public cost attack all point the same way — the open question is whether the acquirers are right or merely early.
What the acquirers are actually underwriting Nvidia threw off $127bn of free cash flow over the trailing twelve months, which means the interesting question was never whether it could afford Hugging Face but what owning Hugging Face is for. Benedict…
3 action items
- ●
Meta Ran the Agent-Substitution Experiment and Cancelled the Layoffs
The best-capitalised attempt to swap headcount for agents failed on outcomes rather than output, and the budget it frees is moving toward supervising agents instead of replacing people.
The agents shipped code and missed the product Reuters, working from scores of internal documents, has Project OT drawn up at Zuckerberg's Hawaii compound in January: cut deeply, hand the removed work to agents, keep a "talent-dense" core supervising virtual…
3 action items
- ●
Nine Months to Tapeout Reopens Custom Silicon — and Tomorrow Prices It
A leading-edge design cycle just compressed below the life of the workload it serves, while a Shanghai listing is about to hand the whole category its first visible public multiple.
Separate the schedule from the marketing Two kinds of claim came off OpenAI's Hot Chips 2026 slides, and only one of them is underwritable. Richard Ho's assertion that the Jalapeño part beats Blackwell on throughput-per-watt and latency, alongside a 56%…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn