Monday, November 16, 2026

Progress is speeding up, but not exploding

Gains per research dollar fell 69 percent as Anthropic's IPO draws 7x demand

Get this in your inbox at 6 a.m. ET, every weekday. Free.

A grey staircase of growing blocks with one mint step, beneath a faint unfilled outline of a vertical curve.

Faster progress, but no explosion yet

The 47-second read

  1. Gains per research dollar fell about 69 percent while spending rose 6.19x, Alex Imas and Luis Garicano calculate → progress is speeding up but getting dearer, so builders must earn back more with each step.
  2. Anthropic's IPO is reportedly about 7x oversubscribed, with up to $90.4B sought, Bloomberg's sources say → the builder layer is being priced for payback on costlier progress, and buyers carry that risk.
  3. Nvidia lost $1.7 trillion in market value in its longest weekly losing streak since 2012, CNBC reports → the chip bottleneck owner is being repriced after Huang's exit, and AI startups face smaller checks.

That's the day. Got more time? Go deeper below.

Go deeper

Today's lead is a careful test of whether AI progress is compounding or merely costly, set against the biggest IPO ever and a shaken chip leader. The question is the same in each: who keeps the value, and what does the price already assume?

The Signal

Faster, yes. An intelligence explosion, not yet.

What happened. Alex Imas and Luis Garicano argue that we are not yet seeing an intelligence explosion. They start from a METR report, built on internal Anthropic data, which put capability growth on well-specified tasks since October 2025 at about 2x the earlier pace (plausible range 1.3x to 3x). Using public S-1 filings, they find annualized R&D spending rose 6.19x, technical staff 3.10x, and capability progress only about 1.91x. That leaves gains per dollar down about 69 percent and gains per technical employee-year down about 38 percent. Their model puts the knowledge-returns parameter at about 0.84 (range roughly 0.66 to 0.94), below the 1 an explosion needs, and they find no strong sign yet that AI is supplying its own research input. They call this an accounting exercise, not proof of cause. Separately, Andrew Curran writes on X that Anthropic reported a similar sudden jump in robotics capability with a model he calls Fable 5.2, and that general models are overtaking specialized training. We have not seen that Anthropic report in today's sources, so treat it as a claim.

Why it matters. This moves two of the seven dashboard readings: the pace of AI research is up, and capability gained per research dollar is down. Imas and Garicano say the speed-up came mostly from inputs growing fast, with the rest looking like a one-time 1.4x to 2.3x jump in research effectiveness, plausibly from coding agents. Holding today's pace would take about a fourfold rise in inputs each year. They also say a sudden burst remains possible once AI does most of the research.

So what for your capital. If each step costs more, the value goes first to whoever sells the inputs that progress consumes, and builders must earn back a larger bill. The authors note that, on Ben Todd's illustrative revenue figures, revenue per index point is rising faster than the cost of progress, so falling gains per dollar do not yet mean falling returns. The exposure is to prices that assume the spending pays back. That is the expensive-progress case, and today's prices look least prepared for it.

Capital Flows

  • Anthropic and some backers are reportedly seeking up to $90.4B through the IPO and a private placement, which Bloomberg's sources say makes it the largest IPO ever → that much demand puts the whole AI trade's hopes and fears on one frontier builder.
  • CNBC describes the IPO as make-or-break for a tech industry heavily invested in the lab and short of good news → concentrated exposure means a weak debut would hit many portfolios at once.
  • Founders and VCs told TechCrunch they expect smaller checks and lower valuations now that Huang has left Nvidia, which put tens of billions into AI startups → startups that relied on a chip supplier as investor lose a source of cheap capital.
  • TSMC plans a $500M fund for chip startups, suppliers and upstream partners, according to Nikkei Asia → the chip bottleneck is paying to create its own next generation of suppliers.
  • Nvidia-backed neocloud Nscale updated its US IPO paperwork showing a widening loss → infrastructure builders are still spending ahead of revenue.
Advertise with The Last Investment

Disruption Watch

Do general models eat the specialists?

Andrew Curran argues on X that as frontier models gain general intelligence, they increasingly beat specialized training, starting with a reported jump in robotics. That is one post relaying a lab's report, and it is too early to call. A second hint comes from Imas and Garicano, who say part of the recent speed-up looks like a one-time gain in research effectiveness, plausibly from coding agents. Software engineering is cognitive work already being repriced. So what: companies selling narrow models, custom training or hourly engineering are exposed if general models keep closing the gap. Whether the decline is faster or slower than their prices assume is unknown today.

Policy and Power

Chips are now a political asset

Scott Bessent urged Nvidia to hire an outside CEO and wants answers within 10 days on whether it pushed "seemingly illicit" chip sales to China, according to Bloomberg. TechCrunch reports new research that China's most advanced models were trained on banned Nvidia chips. Meanwhile the US and Pax Silica allies, including Japan, the UK and India, agreed to draw up a joint list of AI technologies and supply chains to protect from China, ahead of Trump's meeting with Xi in Shenzhen, the Financial Times reports. The bottleneck owner's market now depends on who is allowed to buy.

What Would Prove Us Wrong

The strongest case against the house view is that builders may keep the value this time. Anthropic's IPO is reportedly about 7x oversubscribed, and Imas and Garicano concede that, on Ben Todd's illustrative figures, revenue per index point is rising faster than the cost of progress. Nvidia's slide also shows a bottleneck owner can lose value without the scarcity easing. Heavy demand is not proof of profit, and the revenue figures are illustrative, but the case is not a strawman.

Quick Hits

  • Epoch AI's Ben Cottier estimates Huawei's Horinger data center has about 7% of the compute of SpaceX's Colossus 2 and may stay far behind until at least 2028.
  • A New York Times investigation reports that Howard Lutnick pushed to narrow chip curbs affecting a client of his former firm Cantor Fitzgerald, whose Malaysia data center may serve Tencent.
  • Nvidia closed below $4 trillion in market value for the first time since July 2025, after Huang resigned as CEO following scrutiny over two extramarital affairs.
  • CNBC reports a rotation across chip stocks as Nvidia absorbs the fallout from Huang's resignation.
  • TechCrunch asks whether Huang's troubles hurt the wider AI industry, quoting a source who says much of the economy is riding on Nvidia.

Not investment advice. The Last Investment covers exposures and signals, never recommendations. The author may hold positions in companies mentioned.

Get the 47-second read every weekday

What the singularity means for business and investing, before the market opens.