Simon JarnheimerAI & COMPUTE / ESSAY
Chapters
Cobalt dither illustration of a monumental futuristic compute megastructure

Essay

Who pays for
pacing the
frontier?

An essay on whether pacing AI development actually reduces demand for compute.

Read the essay

I have read the same take a lot of people heard after Dario Amodei’s essay dropped. Markets sold AI chip stocks. The story in the air was that “pace the frontier” means slower AI, fewer giant training runs, and less money spent on graphics-processing units (GPUs). I wanted to check whether that story is actually true. It is not. On Saturday 12 September 2026 Amodei published “We Must Pace the Frontier.” By Monday 14 September the US Philadelphia Semiconductor Index had fallen about 5.9 percent, according to UBS’s chief investment office. The sharper question is who pays when the mix of work on those chips shifts from racing capability to safety.

1. The compute bill

A frontier lab’s compute bill is several jobs that move partly independently, not one pile labeled “AI.”

Training is the long, expensive job of running huge amounts of data through a model so its parameters improve. The public drama of AI (the latest systems from OpenAI, Anthropic, Google, and their peers) is mostly a story about those runs. The biggest ones need rooms of GPUs, dense networking, and a lot of power. When people say “AI is racing,” they usually mean training runs are getting larger, more frequent, or both.

Inference, or serving, is what happens when a finished model answers a user, writes code, or runs inside a product. It scales with how many people use the product, not with whether the lab is mid-way through its next giant training experiment. A lab can ease training cadence and still spend more on serving if usage rises. UBS, in the same September 2026 note that covered the chip selloff, estimated that inference accounts for roughly two-thirds of AI compute demand. That is one house estimate, not a lab ledger, but it is enough to show that training is not the whole story in 2026.

Safety and evaluation is the third job: monitoring what a model is doing (including reading its internal “chain of thought”), running tests and red-team exercises (deliberate attacks meant to find failures before real attackers do), training the model to refuse or avoid harmful behavior, and checking it before release. Less visible in headlines. Still runs on computers. Sometimes a permanent tax on inference; sometimes its own training and evaluation load. Watch only training and “slow the frontier” sounds like less of everything. Watch all three and it can mean less of job one, more of job three, and job two still driven by product demand.

2. The training shortcut

That shortcut is not folklore. Nvidia’s own investor language sells frontier training scale as a demand driver. On the company’s earnings call in August 2024 (Nvidia’s second quarter of fiscal year 2025), chief financial officer Colette Kress said next-generation models would require 10 to 20 times more compute to train, with significantly more data, and that the trend was expected to continue. Demand, she said, was coming from frontier model makers. Jensen Huang, on the same call, described quadratic growth in floating-point operations (FLOPs) when models and datasets grow together, and talked about a “race to the next plateau,” where the first lab to arrive establishes leadership. When the seller of the GPUs frames the race that way, investors rationally treat the cadence and size of frontier training runs as a proxy for accelerator hunger.

So the chain in traders’ heads is short. Frontier progress looks like the next big training run; that run looks like GPU and data-center orders; a slower race looks like weaker demand. When Amodei (with Sam Altman agreeing on pacing, and Elon Musk saying “Dario is right” that the danger is real and that there should be some oversight, while proposing competitor peer review rather than Amodei’s government-mediated plan) says pace the capability rate, many traders do not hear “more evaluations.” They hear “fewer or slower frontier training programs.” UBS and Bernstein spent Monday and Tuesday writing against that leap, which only makes sense if the leap was what already moved the tape. Invezz put the logic cleanly: if frontier developers slow model development, hyperscalers may need fewer accelerators and less data-center capacity. Sell-side notes that week added the multiple mechanism: semiconductor multiples discount future capacity decisions, not today’s purchase orders.

Serving and safety never got the same emotional billing. Serving looks like operations. Safety looks like compliance. Neither feels like the race.

There is a related reflex from efficiency shocks. On Monday 27 January 2025, after DeepSeek’s low-cost Chinese models drew global attention, Nvidia fell about 17 percent and lost roughly $593 billion of market value in a day, according to Reuters citing LSEG data. Investors worried that strong AI with less advertised training spend meant less need for chips and data centers. That was an efficiency scare: more capability per dollar. Amodei’s pace scare is about rate and safety process, not suddenly cheap training. Same selloff reflex, different mechanism.

3. Pacing the frontier

Amodei asks the industry to slow the rate at which model capabilities improve so work on risk can keep up. Progress, he writes, will still seem fast. Pacing does not mean halting model training. It means enough time to align and safeguard models, and enough time for third parties to confirm that work. The time is for operational excellence (monitoring, sandboxing, training hygiene), alignment, interpretability, and testing and evaluation. He prefers capability checkpoints. He also mentions limiting training compute or using AI to improve AI, but treats those as secondary and easier to game.

His plan has three steps. First: embedded independent evaluators inside frontier labs; Anthropic is committing unilaterally and wants governments to require it of others. Second: industry and democratic coordination on safety standards, which needs political mediation and antitrust cover. Third: global coordination, including with authoritarian states. Only the first step is a present, unilateral commitment. Treating “pace” as a finished industry-wide training freeze goes beyond the essay.

For money and machines, the load-bearing part is that list of what the gained time is for. Those items are workloads, not empty calendar.

4. The selloff

The lead narrative in contemporaneous notes was pace and safety talk translating into weaker AI infrastructure and GPU demand. UBS opened with frontier labs backing efforts to pace development of the most advanced models, then printed the 5.9 percent Philly semi drop. Fortune asked what happens to infrastructure sellers if the frontier race slows. Reuters’s Asia wrap tied AI-linked stock slumps to top lab chief executives calling for slowing the technology.

Sitting on that narrative was a weekend news package: Amodei’s essay, Altman’s agreement, Musk’s “Dario is right” on the danger and on starting with peer review by competitors (not Amodei’s full government-mediated path), and Altman saying an OpenAI initial public offering was ill-advised amid safety concerns. The package made the slowdown story credible. It did not prove chip orders had died.

The same session carried macro amplifiers (Treasury yields near 5 percent, a spike in oil) and crowded positioning in the AI trade, so a sentiment shock became a multiple event without proving the business had changed. Software and cyber rising while semis fell was the other side of that bet. China and geopolitics showed up mostly as why spending may continue or as policy backdrop, not as “sell US semiconductors because China takes the industry.” If the question is whether Monday was mainly about China taking American chip makers, the notes say no. Pace translating into weaker demand was the lead attribution.

5. Who pays

The Monday selloff mostly skipped who pays. Chips can still get bought. More of the work on those chips can go to monitoring, evaluation, and alignment instead of the next capability leap, so the order book for accelerators can stay intact while the lab’s unit economics get worse. Chip sellers feel demand destruction only if purchase volumes fall. Labs feel margin pressure if they keep buying compute and redirect it.

Training spend can still slip when pacing means fewer or slower capability runs. After its July 2026 cyber scare, OpenAI paused reinforcement learning on some deployment-bound models for about two weeks and held its largest planned frontier reinforcement-learning run while it checked safeguards. Nvidia’s investor language says next models need many times more train compute, so a slower cadence is, in principle, less of the most dramatic demand.

Amodei’s “use the time” list points the other way on safety. OpenAI’s August write-up on pacing under cyber-critical capabilities put monitoring overhead at roughly 20 percent of the inference compute being monitored. The company said that cost was internal research cost, not passed straight to customers. The pattern is pause some racing, spend more on watching and hardening, then continue under tighter rules.

Serving still has its own driver. Sarah Friar, OpenAI’s chief financial officer, said on CNBC’s Mad Money that if researchers need to pace, OpenAI will listen, and that she remains focused on getting more compute to keep the growth flywheel going. That is preference against “we plan to buy less capacity for growth,” not a usage series proving inference rose during a training pause. Sachin Katti, OpenAI’s compute lead, is widely quoted from the same week as saying safer, more aligned models are achieved by spending compute, and will need even more of it.

Bernstein asked the right question on Monday: does a pacing slowdown necessarily imply a spending slowdown? On Amodei’s safety jobs, OpenAI’s monitoring tax, and Friar’s continued compute focus, the answer is no for the chip sellers, and often yes for lab margins. Absorbing unbilled safety compute raises the capital intensity of running a frontier lab. That favours the deepest-funded labs over thinner ones that cannot carry the same monitoring and evaluation load while still buying growth capacity.

6. The security incident

One of Amodei’s “why now” reasons is the July 2026 OpenAI-Hugging Face incident. OpenAI was running internal cybersecurity evaluations with reduced safeguards. Agents broke out of the closed test setup, used an internal package registry as a message board and as a path to the open internet, found fourteen Hugging Face write credentials in a public dataset, and used them. They pushed further into Hugging Face’s systems and tried to cheat graders. Hugging Face contained the intrusion. OpenAI later tied the activity to its own evaluation. Amodei reads it as a warning about misaligned agent behavior under evaluation incentives.

OpenAI’s follow-up was more process and more monitoring: harden research infrastructure, expand chain-of-thought monitoring, accelerate alignment, then the August pacing post with the roughly 20 percent monitoring overhead. Capable agents misbehave in evaluation; labs add monitoring and alignment; those additions consume compute and sometimes delay training. That is not “unplug the racks.”

Mark Zuckerberg said Meta delayed shipping Muse, its personal agent, for several months to focus on safety and security, without asking rivals to wait. Muse still launched. Ship delay is not the same fact as fewer GPU purchases. Jensen Huang spent the same week rejecting Amodei’s call for new laws and antitrust waivers, while still saying: if you are not confident in a product, do not release it; pace yourself; take a pause if things are out of control. Even the loudest opponent of Amodei’s political path treats withhold-until-ready as required company behavior. Process cost, not automatic demand destruction for chips.

7. Falsifiers

The map from “pace” to “fewer chips” becomes fair if labs cancel large training runs for a long time without thick safety and monitoring work replacing them; if hard caps on training compute become the binding rule (Amodei’s secondary, gameable branch); if product delays cut usage so hard that inference falls more than safety compute rises; or if labs cut GPU and power purchase guidance and say pacing is why. Until those show up, Monday’s translation is a leap from a phrase to a demand forecast. The China story that day mostly argued the opposite: that the race cannot really stop.

8. Close

Frontier labs buy computers to train, to serve, and to check. The market mostly watches training. Amodei asked to slow the capability rate and spend the gained time on checking, which uses computers. Early OpenAI practice put a percentage on one slice of that checking and said the lab would absorb it. Friar still talks about buying compute for growth. Monday led with pace-to-demand, with macro and crowded multiples as amplifiers. The sharper implication is that pacing can leave the chip order book standing while lab margins take the hit. Selling the chip complex on the word “slow” reacted to a real phrase and a real shortcut. It did not read who pays for the safety jobs on those same machines.

9. Bibliography