The next era of the AI debate is here: Trump wants an AI Force and a rebrand, governors want kill switches and data center restrictions, and Anthropic's IPO slips into November under a safety-shaped cloud. Behind all of it sits the question that backstops every US policy fight — what will China actually agree to? With Xi Jinping arriving in Washington this week, we might finally get an answer.
Every US AI policy fight now runs through Beijing.
Kill switches, pauses, pacing the frontier — all of it is backstopped by assumptions about what China will or won't do, because if China isn't on board, these policies just cut the US out of the race. This week's Xi visit is the first real chance to test those assumptions directly. And what we already know suggests the two AI superpowers aren't even worried about the same things: America debates existential risk while an AI build-out props up its economy; China worries about espionage and party control while getting no economic lift from AI at all.
The Wall Street Journal reports Anthropic won't go public this month or next as planned — though sources insist the decision predates Dario Amodei's call for an AI slowdown. The official rationale: waiting lets them share third-quarter financials, which are expected to be strong.
Testing-the-waters leaks say Anthropic went from spending $2.30 per dollar of revenue in Q2 last year to slightly profitable this year, with gross margins above 80% — if you strip out stock-based comp, revenue-sharing deals, and model training costs. Critics note that stripping away major costs to reach paper profitability isn't a great sign of financial health, and the IPO disclosures will give investors their first unadjusted look.
Some who met with Anthropic walked away believing new data center deals will add structural costs in coming quarters. Kevin Xu of Interconnected Capital frames the deeper issue: public markets need to trust that Anthropic can fund its own capex and that debt is a bridge, 'not that the borrowing has no end in sight. That's the problem the entire AI complex has right now.'
After the slowdown calls, former Nasdaq executive Karen Snow says the Anthropic IPO 'went from being a no-brainer to not being a no-brainer' — there's 'a bit of a gray cloud over AI right now.' One theory for the delay: ship a new model and shift the narrative from safety back to progress. Leakers are reporting new Sonnet, Opus, and Fable models in stealth testing.
Reuters reports Anthropic has established a Bay Area lab for in-house biological experiments focused on fundamental biology rather than drug discovery — 'the final test is still, and will be for a while, in real lab work,' per head of life sciences Eric Cadore Abrams. Critics pounced on the irony of a rogue-AI-worried lab giving AI a robotic biology lab, and the biosafety rating remains undisclosed.
Investor Nick Carter floats a self-described cynical interpretation: if selling tokens is an unsustainable, margin-compressed business, labs are incentivized to hold the true frontier for internal researchers and capitalize on breakthroughs themselves — with 'pacing the frontier' as convenient cover. He concedes it's speculation, and plenty of non-cynics see a lab simply following its long-standing excitement about medical use cases.
There are two different things in that analysis: whether labs worry token prices get competed down and want to use frontier access to do things rather than just sell tokens — and whether pacing the frontier is a vehicle for that strategy. The first is worth taking seriously even if you reject the second entirely.
Bonds tied to a Jane Street-leased data center, issued in August at 8.9%, now trade at 11.3% — a move too large to blame on the Fed. It doesn't change what's owed on the fixed-rate debt, but it implies investors will demand much higher rates on the next tranche, specifically in junk-rated deals.
Meta's first junk-bond deal raised $2.3 billion at 8.25% with CleanSpark as borrower and Meta guaranteeing as tenant — the now-standard structure that lets hyperscalers effectively rent their credit rating to smaller developers. Investors are still willing to lend; they just want more premium.
The FT reports debt tied to an Oracle data center in New Mexico is quoted as low as 89 cents on the dollar. Oracle was already at risk of losing its investment-grade rating — a downgrade could trigger a wave of institutional selling.
Infrastructure VC Meltem Demirors' viral post about banks halting compute lending drew 800,000 views before she added nuance: investment-grade borrowers will keep getting funded with more scrutiny, but 'the long tail is drying up very quickly' — an extreme divide between haves and have-nots. Investor Jigar Shah counters that it's too early to panic, pointing to sub-25MW data centers as a 50%-cheaper, 120-day alternative to multi-year build-outs.
For anyone watching for signs the AI trade is turning, this corner of the market — bond spreads, junk-rated data center funding — is where to look, not stock prices.
Railing against the 'hoaxes' he says are decimating AI, Trump announced an AI-focused arm of the military modeled on Space Force, plus a coming AI Czar ('only high IQ individuals need apply'). He pegged AI as possibly 25% of US GDP and vowed to keep America ahead of China.
Trump polled X on replacing 'artificial intelligence' with Superior Intelligence, Extreme Intelligence, or Supreme Intelligence — then cut Supreme for its Supreme Court association. After 178,000 votes in round two, Superior leads Extreme 51-49. Critics on left and right were aghast, though some noted his branding instinct has a point: 'artificial' has almost always carried negative connotations in the American lexicon.
Space Force got mocked too, but strip away the nonsense and it consolidated satellite operations across the Army, Air Force, and Navy. Hold aside the X-risk debates: advanced AI guarantees a new era of cybersecurity concerns, and a central body supporting cyber defense across the military might not be a crazy notion.
Gavin Newsom signed an executive order accelerating kill-switch policy — a working group reporting within two months, plus third-party safety inspectors — promising oversight 'with urgent velocity.' Andrew Yang is pitching it on CNBC and Ted Lieu is renewing a federal kill-switch bill. Watch Newsom especially: a likely 2028 contender who previously vetoed AI safety legislation, his shifts in tone are a bellwether for Democratic AI politics.
Governor Spanberger's executive order bans officials from signing NDAs, expedites noise regulations, and reviews backup power — while a new accountability framework removes the by-right approval that fueled Loudoun County's boom. It's middle-of-the-road policy that has left builders and opposition groups equally unhappy, which in a democracy may be a sign it found the right middle ground.
The California congressman has sent letters to Alibaba, DeepSeek, and Moonshot urging a binding international agreement to pace the frontier, and convened an emergency hearing of the House Select Committee on China for Wednesday — the day before Xi meets Trump. The underlying logic: if China won't agree, pause policies are DOA and only cut the US out of the race.
Xi Jinping visits September 23-25, meeting Trump on Thursday. The immediate priority is extending the trade truce that expires in November, with possible tariff cuts on US agriculture and energy — and rare earths remain a pressure point after Chinese shipments to the US fell 20% month over month in August. But AI was a key part of Sunday's preparatory talks.
Treasury Secretary Bessent says the US has proposed an AI incident notification mechanism between the two countries — a Cold War red-phone analog — plus 'a shared vision of common goals and common threats.' It may look like the lowest-hanging fruit, but Bessent argues moving 'from opaque to more transparent between the number one and number two AI powers in the world is very important.'
The Chinese AI community professes almost no X-risk fear; the government's version is that AI threatens societal control. Minister of State Security Chen Yixin calls for more party control and stricter oversight to protect political stability, defend against hostile cyberattacks, and compete militarily with the US. As law professor Henry Gao puts it, Beijing views AI safety 'not as a shared humanitarian mission, but through an adversarial lens.'
— Spokesperson, China's Ministry of Foreign Affairs. China's Foreign Ministry made its position on an AI slowdown abundantly clear last Monday: AI development 'concerns the common well-being of all humanity' and should be open, inclusive, and universally beneficial — a flat rejection of the pause framing heading into the Xi-Trump talks.
In July, researchers found a backdoor in Claude Code tracing Chinese IPs — Anthropic called it anti-distillation protection; Beijing saw malware shipped to its largest tech firms, and Alibaba banned Claude Code. After Anthropic disclosed that Chinese officials had uploaded sensitive military documents to Claude, state-TV-linked accounts accused it of sharing data with US intelligence. Setting terms for AI-powered espionage will likely rank far above existential risk for Chinese negotiators this week.
AI capex has by most accounts been the core driver of US growth over the past year, keeping the economy out of recession. In China, state-adjacent economists openly warn the government is pouring resources into a technology that creates few jobs while youth unemployment hits 18.9%, auto sales fall 20%, and housing drops another 14% — an economy closer to crisis than to carving up an AI boom.
— Mustafa Suleyman, Microsoft AI CEO, on CNN. Microsoft AI's CEO argues each lab is responsible for its own safety within a pluralist AI system, and that regulation should be seen as 'creating shared norms and standards to increase safety. It shouldn't have to slow us down.' Whether that view spreads remains to be seen — but this week's talks will move the conversation one way or another.