The Shipping Gap
Meta ran the experiment the rest of the industry is still arguing about, and this week the numbers leaked. Its plan to replace teams with small AI-assisted pods produced 220% more code changes year-over-year - and 36% more features that actually reached users. Incidents rose 40%, the time spent cleaning them up rose 70%, and the November wave of cuts was called off. In the same seven days Nvidia posted the biggest quarter in its history and guided to $108 billion, Goldman grew revenue 40% while shrinking headcount, and a federal judge told the Pentagon it may not blacklist an AI lab for refusing to build surveillance tools. Generation is not the bottleneck any more and hasn't been for a while. What reaches users is.
Key Developments
Meta measured what its AI actually shipped, then cancelled the restructuring built on it
Project OT was devised at Zuckerberg's January leadership retreat: collapse 10-to-20-person teams into three-to-five-person AI-assisted pods, thin out middle management, cut some units by as much as 60%. One wave went through, taking about a tenth of staff; the second, scheduled for November, was killed - and a Reuters investigation published August 26, built on internal documents and more than 20 sources, showed why. Code changes to Meta's AI platforms rose 220% year-over-year while features reaching users rose 36%, technical and security incidents rose 40%, the hours spent on them rose 70%, and employee sentiment dropped 19 points. Read those as growth rates, not a yield: code-change growth outran shipped-feature growth roughly six to one, and Meta declined to comment on the figures when Reuters put them to the company. Caveat noted, this is the Klarna Pattern caught one step earlier - Meta stopped before the second round of cuts rather than after the rehiring - and the closest thing yet to a frontier-scale measurement of the gap between what AI produces and what an organization can absorb.
The same week made the opposite case with equally hard numbers. Workday disclosed that more than 5,500 customers now run at least one of its agents, up 35% in a single quarter, with AI driving over a quarter of new contract value; CrowdStrike's AI detection-and-response ARR grew more than 250% sequentially. Narrow agents pointed at a defined job are scaling fine. What Meta cancelled was not an agent product but an org chart premised on one - a bet that you could remove the people who review, integrate and operate the work. The lesson may be about org design, not capability.
Nvidia's best quarter ever landed the same week its biggest customer showed off its own chip
The numbers first: $96.2 billion in revenue, up 106% year-over-year, with data center alone at $89.0 billion, up 117%, at a 75% gross margin. Guidance for next quarter is $108.0 billion give or take 2% - and it explicitly assumes, in Nvidia's own words, that the company books no data center compute revenue from China at all. Then, on the same day, CNBC reported OpenAI unveiling Jalapeno, its first custom inference chip, built with Broadcom. Marvell posted 37% growth the next morning on demand for exactly that kind of custom hyperscaler silicon. Nvidia is selling picks to people who have started forging their own, and still cannot make them fast enough - the guidance is supply-constrained, not demand-constrained.
Custom silicon has been eighteen months away for about four years. Google's TPU took roughly a decade to become genuinely load-bearing, and every hyperscaler alternative so far has ended up absorbing overflow rather than displacing Nvidia - which is why Nvidia can guide to $108 billion, a 12% sequential jump, while writing China down to zero. The read-through that matters is not competitive but accounting: if inference migrates to cheaper custom parts, the useful-life assumptions hyperscalers apply to their existing GPU fleets get harder to defend, which is the mechanism behind P-049.
Goldman grew revenue 40% and got smaller
Record quarter: net revenues of $20.34 billion, up 40% year-over-year, adjusted EPS of $20.98 against a $14.38 consensus, return on equity of 23.5%. Headcount fell 2% quarter-over-quarter. David Solomon described the mechanism without being asked to dress it up - the firm is 'developing marginal levels of revenue production... not growing our human capital footprint quite the same way, but recognizing instead that we need technological capabilities to scale.' That is The Hollowing stated as strategy by a CEO rather than inferred from payroll data: the link between doing more business and hiring more people to do it has been cut, and nobody at the top is pretending otherwise.
Headcount is down 2% for the quarter but still up year-over-year, and CFO Denis Coleman pointedly declined to attach a number to AI's role, calling the efficiency 'an output of the efforts we're undertaking, not a specific target.' A 40% revenue jump at an investment bank is a trading and dealmaking cycle, not a productivity curve - equities and advisory boomed on the AI infrastructure M&A wave itself. Read it as a bank getting paid to finance the buildout, with a modest staffing dividend attached, and the story gets less dramatic.
What the Evidence Moved
-0.15 to 0.56, the full weekly cap, and the analysis supports going further. This claim resolves next month, and the two measures closest to its literal wording both sit below the 20,000-a-month bar: Challenger's AI-attributed cuts average about 16,100 a month across January to July, and Goldman's own tracker has been revised from roughly 16,000 net a month in April to about 11,000 by June. That improvement is not displacement slowing - Goldman's substitution column still runs near 25,000 a month, offset by around 9,000 a month of data-center construction hiring that has nothing to do with AI sparing anyone. The deeper problem is that no authoritative monthly series for 'AI-exposed industries' exists, and the three proxies in use diverge enough to resolve this either way. The September note must name one metric rather than blending them.
+0.02 to 0.93. Workday disclosed on August 27 that over 5,500 customers run at least one of its agents, up more than 35% in a quarter, with AI behind more than a quarter of new contract value; CrowdStrike put Falcon Flex ARR above $2.29B (+101%) with AI detection-and-response ARR up over 250% sequentially; Okta framed its quarter around agent identity, the plumbing you only build once agents are actually in production. Held small: these are management-selected metrics and 'at least one agent' is a softer bar than Gartner's 40%-of-applications wording.
-0.04 to 0.18. Six months on, Block is still the only company to have cut 40% of its workforce in one restructuring while citing AI. Across roughly 19 large firms that blamed AI for 2026 layoffs, the cuts cluster at 10-20%. The nearest analog is Jamie Dimon saying AI cut JPMorgan headcount 30-40% in certain divisions - per-division, not company-wide, with staff redeployed rather than replaced. Meta's Project OT contemplated up to 60% on some teams and was cancelled. Non-replication at this sample size is real evidence; Block's own case stands unreversed, which keeps this above 0.15.
-0.03 to 0.47. Meta cancelled the reorganization premised on agents doing the work after its own metrics showed code changes up 220% against features shipped up 36%, incidents up 40% and incident time up 70%. Read the 220% carefully - it measures changes to Meta's AI platforms, not the share of Meta's code written by AI, which is what this claim resolves on. But a company that pulls the November cut wave is not a company about to announce that agents write most of its code.
Company Impact
Sources
- NVIDIA - Q2 FY2027 financial results ($96.2B revenue, $108.0B Q3 outlook, no China DC compute assumed)
- Engadget - Meta abandoned an AI-focused restructuring plan (carrying the Reuters investigation)
- Sunday Times - Special report on Project OT (same Reuters investigation, syndicated; Meta declined to comment on the internal figures)
- Al Jazeera - US judge blocks Pentagon blacklisting of Anthropic
- Workday - Q2 FY2027 results (AI >25% of new ACV, 5,500+ agent customers)
- Goldman Sachs Q2 2026 earnings call transcript (net revenues $20.34B, headcount -2% QoQ)
- European Commission - AI Office (enforcement powers, no listed actions)
- Tech Policy Press - Data center discontent drives state legislation surge
- Artificial Analysis - Kimi K3 Intelligence Index placement
- MarkTechPost - Z.AI releases GLM-5.3-Flash (320B-A18B MoE, MIT license)
- Indeed Hiring Lab - US labor market snapshot, August 2026
- Challenger, Gray & Christmas - July 2026 report (AI leads cut reasons fifth straight month)
- Cursor - Joining SpaceX
- Equinix - Q2 2026 results and raised long-term outlook
- Bloomberg via TNW - JPMorgan leads $5B debt package for Volta AI buildout
- CNBC - OpenAI unveils Jalapeno, its first custom AI inference chip with Broadcom
Next issue drops Monday
Subscribe to get the briefing before the market opens.