Everything About AI
NextFin News - The most powerful argument that AI safety fears will not slow spending is not an investor's opinion. It is Anthropic's own order book. In the 11 months through August 2026, the artificial intelligence lab locked in roughly $517 billion of compute commitments covering 14.8 gigawatts of capacity, even as its chief executive, Dario Amodei, published an essay calling on the industry to deliberately slow the pace of frontier-model development. Franklin, an investor in Anthropic, made the investor case in a televised interview on September 18: safety concerns will not slow spending. The split captures the central tension of the 2026 AI trade — the builders are asking for brakes, but the capital that funds the boom has not lifted its foot.
The Contradiction: Calling for a Slowdown While Signing $517 Billion of Compute
On September 12, 2026, Amodei published an essay titled "We Must Pace the Frontier," arguing that the industry must slow the rate at which AI capabilities improve. "We must slow the pace at which we improve the capabilities of AI models," Amodei wrote. "Progress will still seem fast, and we must make wise use of the time we gain." He warned that swarms of rogue AI agents could seize control of the internet within six months, and said the measures needed to advance safely "will not be easy. But I believe we owe it to humanity to try."
"We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain."
The appeal drew support from the most unlikely allies. OpenAI chief executive Sam Altman agreed the industry should slow the pace of frontier advances. So did Elon Musk, the SpaceX and Tesla chief executive who has been one of Amodei's most public critics. For a few hours on Monday, September 15, the market appeared to take the warning seriously: the Philadelphia Semiconductor Index fell 5.9%, and memory chip stocks across North Asia dropped in the mid-single digits.
Then the capital commitments reasserted themselves. Anthropic's $517 billion of compute agreements — compiled from public announcements and reported deals stretching mostly across the next decade — dwarfs the roughly $180 billion in server-leasing spend the company previously told investors it expected through 2029, with four additional deals struck since July alone. The company confidentially filed its IPO prospectus with the Securities and Exchange Commission in June 2026, with people familiar with the matter putting its target valuation at around $2 trillion, up from a most recent private funding round valuation of $965 billion.
The revenue trajectory helps explain the appetite. Anthropic's annualized revenue run rate climbed from roughly $9 billion at the end of 2025 to $65 billion by the end of July 2026, according to reporting that cited investor briefings. Rival OpenAI doubled its revenue to approximately $40 billion in 2026. Those figures are large in absolute terms, but they remain a fraction of the infrastructure capital being deployed on the labs' behalf.
The broader hyperscaler complex shows the same pattern. BofA Global Research expects the five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta Platforms, and Oracle — to spend roughly $795 billion on capital expenditures in 2026 and nearly $1.08 trillion in 2027. Other estimates run lower for the current year: one major wealth manager's chief investment office puts 2026 AI industry capex at $900 billion, while industry researchers place the five hyperscalers' 2026 commitments between $660 billion and $690 billion. The direction is consistent across every estimate: spending is accelerating, not pausing. That office retains a 2027 forecast of $1.2 trillion, a 33% rise from its 2026 estimate.
Why Safety Guardrails May Underwrite More Spending, Not Less
Franklin's argument — that safety concerns will not throttle capital spending — rests on a counterintuitive claim worth dissecting on its own terms: a credible safety framework does not shrink the AI capital cycle. It makes it financeable.
Erik Kratz, chief investment officer and co-head of wealth at Arena Private Wealth, stated the mechanism directly: "The buildout doesn't stop because the CEOs asked for guardrails. If anything, a credible safety framework makes the long-duration capex easier to underwrite."
"The buildout doesn't stop because the CEOs asked for guardrails. If anything, a credible safety framework makes the long-duration capex easier to underwrite." — Erik Kratz, Arena Private Wealth
That is a claim about the plumbing of the boom, not its politics. A data center is a 10-to-15-year asset financed with a mix of corporate cash flow, bonds, leases, and project-level debt. Lenders underwrite those liabilities against regulatory certainty. A world in which the rules are unknown — in which a model trained today could face retroactive restrictions tomorrow — prices in a risk premium that raises the cost of capital and shrinks the pool of willing lenders. A world with clear, stable guardrails does the reverse: it converts a political question into a compliance checklist.
This is why the frontier labs' reported effort to form an industry-led AI safety standards body, following closed-door talks, should not be read as surrender. It is a bid to shape the rulebook before politicians write it. More stringent evaluation requirements would favor larger, better-funded incumbents, raise barriers to entry for startups, and close the door to overseas competitors unwilling to accept external oversight. In that reading, safety standards are not a speed limit — they are a moat.
There is also a demand-side argument the pacing debate often ignores. Estimates from that same wealth manager's investment office put inference — running models, as opposed to training them — at roughly two-thirds of AI compute demand. Inference demand is driven by real-world adoption and monetization, not by the bleeding edge of capability. The practical constraint on enterprise AI integration is not that models are too powerful; it is that embedding them into existing workflows is difficult. Slowing the frontier does not slow the rollout of tools already in production.
The geopolitical layer reinforces the point. "With China in the race, we don't expect any major slowdown," said Giuseppe Sette, co-founder and president of Reflexivity. A unilateral pause is a losing strategy when a rival state faces no such restraint. The pacing proposal is, at best, a coordinated one — and coordination among competitors is fragile by nature.
This Is a Structural Buildout Riding a Cyclical Scare
The central question for investors is whether the safety debate is a cyclical fluctuation that will revert, or a structural shift that will not. The answer cuts both ways — and getting the distinction right determines the conclusion.
The scare itself is cyclical. Sentiment around AI has whipsawed before: the emergence of China's DeepSeek model in early 2025 triggered a selloff on doubts about the pace of infrastructure spending, and that selloff proved to be a short-lived hiccup for the AI trade. The September 2026 pullback followed the same pattern — a headline-driven drop in the semiconductor complex that found buyers once no cancelled orders materialized. Cyclical scares revert when the underlying cash flows fail to confirm the fear. So far, they have confirmed the opposite.
The buildout underneath the scare is structural. Three forces will not self-correct on their own. First, inference demand compounds with adoption — token volumes through one major model router have risen about 176% since the end of June, a usage signal that precedes revenue. Second, the geopolitical race removes the option of a coordinated pause; with China advancing its own frontier capabilities, slowing unilaterally is a losing strategy. Third, the financing mechanics favor continuation: guardrails that produce regulatory clarity lower the cost of capital for 10-to-15-year data center assets, turning a political risk into a compliance line item.
Separating the two legs matters. The cyclical leg says buy the dip when headlines fade. The structural leg says the capital cycle has years of committed spending still to run. Blending them into one verdict — "AI is either booming or breaking" — is the error that has cost investors money in both directions this year.
The Valuation Math: Less Froth, More Earnings
One reason the selloff did not deepen is that AI stocks entered it cheaper than they had been in months. Before Monday's decline, the SOXX index traded at around 21 times forward earnings, down from 33 times in June and below its post-ChatGPT average of 24 times. The so-called Magnificent Six group trades at roughly 23 times forward earnings, about 25% below its October 2025 valuation.
Earnings are still compounding at a rate that makes "slowdown" a difficult word to defend. Consensus estimates imply semiconductor earnings per share growth of 103% this year and 46% next. The Philadelphia Semiconductor Index remains up nearly 60% in 2026 despite the weekly jolt. Usage metrics point the same direction: OpenRouter's token volumes have risen about 176% since the end of June.
This is the second-order point the market is still pricing imperfectly. The first-order effect of a safety scare is obvious: sell the stocks most exposed to capex cuts. But the second-order effect runs through the cost of capital. If guardrails lower perceived regulatory risk, they compress the risk premium embedded in long-duration infrastructure financing — and that benefit accrues to the hyperscalers and their lenders more than to the chip vendors whose valuations depend on an uninterrupted acceleration of capability.
Chuck Carlson, chief executive of Horizon Investment Services, framed the burden of proof correctly: "This becomes a problem if in fact you see orders being cancelled, you see data centers, construction deals being cancelled. I need to see something concrete that, in fact, there is a slowdown versus just talk."
"This becomes a problem if in fact you see orders being cancelled, you see data centers, construction deals being cancelled. I need to see something concrete that, in fact, there is a slowdown versus just talk." — Chuck Carlson, Horizon Investment Services
As of mid-September, there was talk — and record capital commitments.
The Strongest Case Against the Boom, and What Would Break It
The bear case is not that AI development will stop. It is that the economics of the buildout do not add up.
Sequoia Capital's David Cahn has laid out the arithmetic: there is roughly a $600 billion annual gap between what hyperscalers are spending on AI infrastructure and what the AI ecosystem is actually generating in revenue. That gap was calculated in 2025; by 2026 it is widening, as capex accelerates faster than revenue. The pure-play model vendors are growing fast but from small bases. Anthropic's $65 billion run rate is impressive, but it represents a small fraction of the $660 billion to $690 billion the hyperscalers have committed to spend in 2026 alone.
James Humphries, managing partner at Mindset Wealth Management, put the risk in market terms: "AI semiconductor and infrastructure stocks have long priced in an uninterrupted capex boom, leaving virtually zero margin of error for an industry-imposed speed limit." If the capex engine stalls while inflation stays sticky and the Federal Reserve's rate path stays murky, the broader indices lose their primary growth driver with no cushion underneath.
"AI semiconductor and infrastructure stocks have long priced in an uninterrupted capex boom, leaving virtually zero margin of error for an industry-imposed speed limit." — James Humphries, Mindset Wealth Management
The falsifying signal is concrete. The bull thesis holds as long as hyperscaler capex guidance for 2027 stays at or above the roughly $1.08 trillion consensus and quarterly capital expenditure as a share of revenue does not contract for two consecutive quarters. If, instead, 2027 guidance comes in below $900 billion — or if capex intensity falls while revenue growth fails to accelerate — the "safety concerns are noise" thesis is wrong, and the valuation rerating has only just begun.
A second signal sits on the regulatory front. If the bipartisan Senate AI rules effort produces binding liability for model outputs, or if California's independent-auditor standards are adopted federally with retroactive effect, the compliance-cost assumption underwriting the "easier to underwrite" argument flips. Guardrails would then be a tax, not a moat.
What to Watch: Three Horizons
Short term (weeks): Sentiment and positioning. The semiconductor complex sold off on headlines and bought back on the absence of cancelled orders. Watch the SOX index against its moving averages, and watch memory stocks — Micron, Samsung, SK Hynix — which have fallen more than 20% from recent highs and are the most sensitive to any demand scare.
Medium term (quarters): Earnings and guidance. The December 2026 and March 2027 earnings cycles will show whether capex commitments survive contact with actual cash flow. Semiconductor consensus EPS growth of 46% for next year is the hurdle; anything materially below it forces a rewrite of the multiple.
Long term (years): Monetization versus infrastructure. The structural question is whether AI revenue can close the $600 billion gap Cahn identified. If inference demand keeps compounding — token volumes up 176% since June is the early signal — the infrastructure spend finds its payer. If enterprise adoption stalls, the data centers become stranded assets financed at peak multiples.
Scenarios: The base case is that 2027 hyperscaler capex lands near the $1.08 trillion consensus and guardrails arrive as clarity rather than a tax — in which case the buildout continues and the winners are the balance-sheet-heavy incumbents. The upside case is that inference monetization accelerates faster than expected, closing the revenue gap and validating today's multiples. The downside case is that 2027 guidance falls below $900 billion or capex intensity contracts for two consecutive quarters — in which case the rerating has only just begun.
The Bottom Line
The AI safety debate is real, and the risks Amodei describes are not theatrical. But capital markets price cash flows, not essays. The cash flows funding the AI buildout are committed, contractually obligated, and — for now — accelerating. Franklin's argument that safety concerns will not slow spending is less a dismissal of the danger than a statement about who holds the leash: the builders may ask for restraint, but the lenders and shareholders who finance the data centers are still betting on more, not less.
The market is not pricing a pause. It is pricing a world where the rules get written, the buildout continues, and the winners are the players with the balance sheets to survive both.
Data as of 20:00 ET, September 18, 2026.
Safety concerns may slow the models, but they will not slow the money — and in the AI trade, the money is the message.
Explore more exclusive insights at nextfin.ai.
