The Week the Story Stopped Being Enough
On September 10, 2026, Adobe reported quarterly earnings after the market closed, choosing to emphasize growth in its user base rather than the performance of its paid products.1 Investors did not wait for the full picture. The fourth-quarter guidance that followed fell short of analysts' expectations, and Adobe's stock dropped nearly 5% in after-hours trading before recovering by Friday morning.2 It is a small, specific data point, but it captures something larger happening across the AI sector this month: companies that once got credit simply for having an AI story are now being asked to show the number that goes with it.
Alphabet's Own Numbers Tell a Related Story
The same tension is visible at a much larger scale. Alphabet's second-quarter 2026 results, reported in July, showed sales of $119.80 billion and net income of $112.19 billion — a single quarter's revenue on a scale that outproduces the yearly economic output of many countries.3 Alongside those results, Alphabet filed to shelf $69.24 billion in new Class C stock, kept its dividend in place, and is contending with a newly certified class-action lawsuit that has not yet reached a settlement or judgment.3 The detail drawing the most attention: Alphabet has paused share buybacks under its 2025 program and is spending so heavily on AI infrastructure that free cash flow has turned negative.3 Via News notes that this figure comes from a source whose claims have checked out about 57% of the time in our historical fidelity checks — worth including, but a reminder that even business-wire financial reporting isn't automatically gospel. It's also worth noting Alphabet is not just a bystander in this story: as the next section shows, it's one of the companies now losing safety researchers over the same technology driving that spending.
Who Is Leaving, and Why
Two AI researchers left Anthropic and Google on September 10, 2026, both citing safety concerns about the systems they were building — arriving one day after Anthropic researcher Jacob Coxon's own departure went viral.4 Elon Musk, whose SpaceX and OpenAI both compete directly with Anthropic,10 dismissed the departures as a "setup" and a "psyop."4 The pattern isn't confined to those two companies. Brad Lightcap, OpenAI's special projects lead and former chief operating officer, also announced his departure; The Verge reported that he "added that he'd still be around for the next few weeks."5 Three separate exits, at three of the most closely watched AI labs, inside the same news cycle as guidance misses and cash-flow warnings — that overlap is the story as much as any single departure.
Regulators Move, and So Do the Companies' Own Disclosures
California enacted a law on September 10, 2026, banning infinite scroll and autoplay for users under 16 and introducing the first rules in the country specifically covering AI companion chatbots.6 The same day, Anthropic published its own account of how Claude was being misused: the company said it had identified and blocked five potential plots to build biological weapons and six attempts to use its AI to write software for conventional weapons, and disclosed separately that governments had used Claude for surveillance and that Russia-linked hackers had targeted it.7 Publishing that kind of report voluntarily is unusual, and it lands in the same week as a state law aimed at a different AI harm entirely — evidence that scrutiny is arriving from regulators and from the companies themselves at the same time.
The underlying argument over how much of this risk is real isn't new, and it isn't confined to Silicon Valley. Discussing AI policy earlier this year, John MacIntyre put it this way: "On one side are the doommongers who think AI is going to end civilisation. On the other are the blind optimists who argue any regulation is simply an attempt to stifle innovation. The truth lies somewhere in the middle."8 That comment appeared in an announcement whose broader claims have held up only about 31% of the time in our checks, so treat the surrounding context loosely — but the quote itself is attributed directly to MacIntyre, and it frames the current moment fairly: California's law and Anthropic's disclosure both sit in that middle ground, treating AI as neither harmless nor something to be shut down outright.
Who Sells to Governments, and Who Sells Protection
Behind the earnings and the departures sits a less visible layer: who supplies the computing power, and who has direct relationships with governments. Anthropic buys chips from both NVIDIA and Broadcom and holds a contract with the Pentagon10 — putting it, alongside Palantir Technologies, which is contracted by the U.S. government and built the Gotham data-analysis platform for exactly that kind of client,10 inside a small group of AI companies with direct federal ties. That same government exposure is likely part of why OpenAI chief executive Sam Altman spent September 10 pitching his company's cyber-defense capabilities directly to power and utility companies, following reports of AI-enabled attacks on critical infrastructure.9 Selling protection against AI-enabled threats while your own safety researchers resign over AI's risks is a position more than one of these companies now occupies at once.
What to Watch
Four threads from this week are worth tracking rather than treating as settled: whether Adobe's stock recovery holds beyond the initial rebound;2 whether Alphabet's negative free cash flow persists into its next quarterly report;3 whether other AI labs follow Anthropic's example of publishing a misuse report rather than staying silent;7 and whether California's under-16 rules become a template other states adopt.6 None of these resolve the underlying argument — MacIntyre's point that the truth sits between doom and blind optimism — but each is a concrete, checkable data point rather than a narrative, which is exactly the shift this week's news suggests investors are now demanding.
