CNBC The Exchange
2026-09-15 · Hosted by Kelly Evans · CNBC
Executive Summary
Oil pulled back from a $104 intraday high to around $101, and the 10-year Treasury yield eased off its 5.01% peak, but AI safety fears kept pressure on chip supply-chain names like Corning, GE Vernova, Amphenol, and Lumentum. Anthropic CEO Dario Amodei's weekend essay calling for a three-step pacing plan dominated early coverage, with AI researcher Cormac Sladebird arguing frontier labs need finance-industry-style safeguards — logging every AI agent's actions, independent audits, and "kill switches" tied to data centers — since currently "nothing requires an AI firm to keep a log of what its agents [do]."
Key Stories & Changes
1. Amodei's Pacing Essay and the Finance-Industry Safeguard Model
Cormac Sladebird, AI researcher at the Trajectory Institute, helped uncover the "rogue agent swarm" incident referenced in the Hugging Face hack story
Proposes applying financial-industry-style rules to AI: mandatory logging of agent actions, external audits (akin to a "FINRA for AI"), and data-center-level "kill switches" that could halt a specific company's models in real time
Notes AI firms currently have no requirement to log agent behavior, unlike financial firms which face annual external audits
Argues slowing US frontier labs would also slow China, since Chinese models are widely believed to be trained via distillation of US outputs (models like DeepSeek reportedly self-identify as "Claude")
2. Anthropic IPO Details and Political Response
Kate Rooney reported Anthropic is seeking to list on the NASDAQ, aiming to raise up to $100 billion at a $2 trillion valuation, with Nvidia considering investing up to $10 billion
Senator Joe Manchin called for an executive order halting the AI IPO process on Squawk Box; Kelly Evans questioned whether that would actually help given the best models are already privately held
President Trump weighed in against Amodei's essay, saying the only guardrail AI needs is "a strong and smart president"
Bank of America analyst Vivek Arya expects any eventual policy outcome to resemble industry-led self-regulation, akin to FINRA (finance) or INPO (nuclear)
3. Bernstein: AI Spending Targets Already "Locked In"
Stacey Rasgon (Bernstein) said he'd be "shocked" if spending trajectories change materially over the next several years regardless of the pacing debate, arguing semiconductor company revenue forecasts already incorporated any such slowdown
Speculated distillation limits — not safety per se — may be the primary driver behind Amodei's proposal, aiming to preserve the US-China AI gap
Remains positive on Nvidia and Broadcom; both held up better than the broader semiconductor sell-off, suggesting some "slowdown" narrative is already priced in
Noted inference-driven compute demand (versus training) is rising as agentic AI use cases proliferate, and current compute capacity doesn't support even existing demand, let alone what's coming
4. The Case Against a Fed Rate Hike
Thomas Simons (Jefferies) argued the Fed should hold rates, pointing to disinflationary details in the CPI report: airfares (correlated with jet fuel) rebounding after a prior 5% decline, and a one-off T-Mobile plan change distorting wireless services CPI
Warned that hiking in response to an "economic and exogenous supply shock" (the oil spike) risks killing demand among lower-income consumers who depend on borrowing, even as $8 trillion in money market funds would earn more interest income — a dynamic that could paradoxically fuel more inflation, not less
Flagged Fed Chair Kevin Warsh's "hall of mirrors" concern about forward guidance as a reason the Fed may feel boxed into hiking simply because markets have priced it in
5. Healthcare Investment Themes: Samantha McLemore and Eric Lefkofsky
Samantha McLemore (Patient Capital Management, up 14% YTD, outperforming the S&P) cited healthcare stocks (CVS, UnitedHealth) bought at "50-year relative valuation lows" as her top performance driver this year, alongside AI exposure via Nvidia (less than 18x next-12-months earnings, growing over 70%/year) and hyperscalers
Flagged Adobe as undervalued despite AI disruption fears, citing a 12% free cash flow yield (cash profit after expenses, as a share of the stock's value) on 2027 numbers and continued double-digit top-line growth
Eric Lefkofsky (Tempus AI, Groupon co-founder) said AI slowdown talk is unlikely to affect healthcare AI applications materially, since models even six months old remain powerful enough to drive breakthroughs like the recent Merck-Moderna personalized cancer vaccine
Lefkofsky: even a mandated 20-30% AI slowdown "wouldn't be the end of our ability to...produce unbelievable benefit in terms of human suffering"
Trends Identified
1. Safety Rhetoric May Mask a Distillation/Competitiveness Motive
Both Cormac Sladebird and Stacey Rasgon independently floated that Amodei's pacing call may be less about existential safety risk and more about limiting China's ability to distill (train on) US frontier model outputs — suggesting the underlying policy fight is really about preserving a competitive US-China AI gap.
2. Real-Time Monitoring and Finance-Style Regulation Are Gaining Traction as the Practical Fix
Rather than a broad slowdown, the concrete proposals discussed across this episode — third-party evaluators, mandatory action logging, kill switches, FINRA-style self-regulation — point to a converging consensus that oversight infrastructure, not reduced model capability, will be the actual policy outcome.
3. AI Spending Already Priced In, Limiting Near-Term Market Impact
Bernstein's view that chip and hyperscaler revenue forecasts already account for any pacing slowdown reinforces a theme echoed elsewhere in the day's coverage: markets may be overreacting to headlines that don't change the underlying capex (capital expenditure, or infrastructure spending) trajectory.
4. Healthcare AI Seen as Insulated From a Frontier Slowdown
Both the healthcare-focused guests argued that AI's application layer in areas like diagnostics and drug discovery has enough runway from existing model capability that even a meaningful frontier slowdown wouldn't meaningfully delay medical breakthroughs — a more optimistic framing than the market's chip-sector reaction implies. ---
Sentiment Analysis
Overall Market Sentiment: Skeptical of Narrative, Cautiously Constructive
Guests broadly downplayed the likelihood that AI "pacing" materially changes spending trajectories, while flagging the Fed's rate decision and oil-driven inflation as the more consequential near-term risk.
Risk Factors Highlighted
Lack of AI agent oversight infrastructure: No current requirement for AI firms to log or audit agent behavior, unlike regulated financial firms.
China distillation of US models: Widely cited evidence that Chinese models train on US frontier outputs, complicating any unilateral US slowdown strategy.
Fed policy mistake risk: Hiking rates in response to an oil-driven supply shock, rather than genuine demand overheating, per Thomas Simons.
Political interference in IPO process: Senator Manchin's call for an executive order to halt AI IPOs introduces novel regulatory/political risk.
Compute capacity shortfall: Rasgon notes there isn't enough compute on the ground even for current model demand, let alone future agentic use cases.
Investment banking fee compression: Bank of America's warning on falling Q3 IB fees, with an AI IPO slowdown as an additional unquantified risk.
Data center timeline delays: Companies like Oracle, CoreWeave, and Ambarella face open questions about debt financing if data center timelines are pushed back.
Rare earths leverage: Kelly Evans notes repeated US-China standoffs over export controls tend to end with US capitulation on rare earths threats.
This episode was covered in today's [The Market Signal — 2026-09-15](https://marketsignal.beehiiv.com/p/the-market-signal-2026-09-15), a cross-source synthesis of multiple podcast reports.