Anthropic has filed to go public, and the most interesting number in the filing is one two outlets can’t agree on. Elsewhere: a private equity economist says AI has paid off for almost nobody outside Big Tech, and the PHP emergency shows what a leaky disclosure process costs.

1. Anthropic’s IPO pitch: buy in, but don’t expect a say

Anthropic’s S-1 asks prospective shareholders not to think of the purchase “in purely transactional terms.” The governance section makes that easy to follow, since the structure leaves little else to do with the stock. Seven co-founders, grouped as the “Founder LLC,” will hold 50.1% of voting power, and the Long-Term Benefit Trust will appoint most of the board. TechCrunch notes that voting power won’t track the founders’ economic ownership.

The financials are messier than the headlines. Both WIRED and TechCrunch put first-half 2026 revenue above $16 billion, but they give different prior-year comparisons ($4.5 billion for 2025 versus $1.2 billion for the same period). They also disagree on the net loss: WIRED says $80.7 million, TechCrunch says $80.7 billion, which it attributes largely to noncash accounting adjustments on convertible debt. I wouldn’t lean on either until the filing itself is checked. TechCrunch’s operating loss of $693 million, down from $2.1 billion, is the steadier figure.

The argument I’d make: the safety mission is doing double duty as a control mechanism. Investors get exposure to $16 billion in growth and a risk list that includes “catastrophic or existential risks to humanity” and the government’s supply-chain-risk designation. They get very little leverage over any of it. The S-1 still doesn’t say what Anthropic hopes to raise.

TechCrunch · WIRED

2. Apollo’s economist finds no AI payoff outside the Magnificent Seven

Apollo chief economist Torsten Slok looked at the S&P 500 minus Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. In that “S&P 493,” he writes, profit margins and revenue per worker look flat, with no sign that AI is improving productivity. His summary is blunt: “zero financial payoff to corporate America outside of tech.“

The caveat is that this is one economist’s note on aggregate numbers, as reported by TechCrunch. Flat company-wide margins can hide real gains in individual teams or workflows. They can also reflect adoption that is still early. Slok doesn’t seem to accept the “just lagging” reading, though. His case is that cheap intelligence doesn’t automatically become higher revenue, because business runs on judgment, accountability, trust and unwritten rules.

That should worry anyone selling AI to enterprises on a productivity pitch. Buyers will increasingly ask for evidence in margins and revenue per worker, and a note like this gives skeptical CFOs something to cite. TechCrunch adds that other recent studies have also found no significant productivity gains yet.

The Register · TechCrunch

3. PHP2Shell’s patch race started before the patch

The PHP2Shell timeline is the real story. According to The Register’s Kettle podcast, an independent researcher found the memory-safety flaws in August, and coordinated disclosure with the PHP Foundation began. It frayed over a weekend, then a video purporting to show an exploit leaked. The Foundation warned of emergency patches late Monday and shipped them Wednesday. AWS then said it had seen exploitation attempts at least four days before the patches.

The scale is the other problem. The podcast says the bugs affect releases back to about PHP 5.4, and that roughly 70 percent of websites run PHP, a figure Lyons describes as consensus rather than a hard count. As of Friday, the Foundation and experts she spoke to had seen attempts but no confirmed successful exploitation. Old servers with no direct upgrade path remain exposed. And a third flaw, which the Foundation says needs an unlikely configuration, will wait for a regular release.

The panel’s argument is that disclosure timing and legacy sprawl are colliding with AI-assisted bug hunting. Liam Proven adds that many sites run PHP without knowing it, because their CMS does. Only one outlet’s podcast is the source here, so treat the framing as commentary.

The Register · The Register

4. Britain wants AI agents guarding critical infrastructure

The Register reports that Whitehall wants AI agents guarding critical infrastructure, as part of a wider £600M spending spree. That’s all the available text says, so the headline claim stays modest: the £600M is the broader package, not necessarily the agent budget.

Still, the framing is the point. The UK is pitching AI as a defense against threats that AI is helping create. That implies a bet on machine-speed response, and it raises an obvious procurement question: who builds, audits and controls agents with access to critical systems? The source gives no answers yet.

The Register

5. Pat Gelsinger backs another attempt at the memory wall

The Register has an exclusive on Nagare Labs, the startup former Intel CEO Pat Gelsinger is backing. Its Z2RAM promises SRAM-like bandwidth, DRAM-like density and NAND-like manufacturing. The Register itself warns that the road ahead is “arduous and littered with the ruins of failed memory vendors.“

That warning is the useful part. Memory is the long-standing bottleneck for AI hardware, so a credible fix would matter. But the available text is only a teaser, with no benchmarks, no silicon and no timeline. A famous backer is not a working product, and a promise to beat three technologies at once is a lot to take on faith.

The Register