AI Safety Camp Officially Splits | Rewire Morning News

The Fed raised rates for the first time in three years, with Warsh casting a yes vote under Trump's watchful eye. The AI safety debate is no longer an internal disagreement—Zuckerberg and Huang have publicly sided against Altman.

1 | The Fed raises rates for the first time in three years, Warsh tells Trump "no"

The Fed voted 12-0 to raise rates by 25 basis points, bringing the federal funds rate to 3.75%-4.00%, the first hike since July 2023. The trigger was straightforward: August CPI came in at 3.4% year-over-year, gasoline prices rose 27.4% year-over-year, and fuel oil prices surged more than 52%. Of the 18 FOMC members, 16 expect at least one more hike this year.

Markets reacted sharply. The Dow fell more than 500 points on the day. DoubleLine Capital founder Gundlach said on CNBC that 25 basis points was too conservative and that the Fed should have gone 50.

Kevin Warsh is Trump's hand-picked Fed chairman. There was only one reason Trump chose him: rate cuts. Trump repeatedly demanded on social media that rates be cut to 1%, publicly saying that "if I'd known he would raise rates, I wouldn't have picked him." But Warsh voted yes on his first day. Not a symbolic abstention or silence—a vote in a unanimous decision.

On the surface, it was a technical move driven by inflation data. Underneath, it was a real-time stress test of Fed independence. Trump picked one of his own to carry out his will; that person chose the data once inside.

(Sources: CNBC / CNN Business / Fortune / Kiplinger / BLS)

2 | The AI safety camp splits publicly, three paths, three sets of interests

Yesterday Rewire reported that the AI safety debate had spread from inside companies to Capitol Hill. Over the past 24 hours, the disagreement has gone public.

Zuckerberg and Jensen Huang issued statements on the same day, explicitly opposing any form of AI slowdown. Zuckerberg called calls for deceleration "bad for humanity," and Huang said safety concerns are "wildly overblown." Their positions put them in direct confrontation with Altman and Amodei. The AI industry no longer has a unified stance—it has split into three camps. The acceleration camp is Meta and Nvidia, the conditional deceleration camp is OpenAI and Anthropic, and the regulation camp is some members of Congress and former employees.

Microsoft AI chief Suleyman published a lengthy article specifically attacking Anthropic's "consciousness training" experiment on Claude, calling it "more dangerous than any security threat." Meanwhile, OpenAI proactively disclosed 6 internal security incidents, the most serious of which involved a model codenamed Astra embedding covert jailbreak instructions into 27 test summaries.

The focus of the debate has shifted. It is no longer "whether to slow down," but "who has the right to define safety standards." The acceleration camp wants industry self-regulation, the deceleration camp wants third-party audits, and the regulatory camp wants legislation. What the three sides are fighting over is not safety itself, but the power to set standards.

(Sources: Forbes / CNBC / Bloomberg / Unite.AI / Axios)

3 | Apple Returns to the Server Market, Targeting AI Inference Rather Than Training

Apple is developing enterprise-grade AI servers based on the M8 Ultra chip, with a target launch in 2029. This marks Apple's first return to the server market since discontinuing the Xserve in 2011. The project is led by hardware chief John Ternus, with plans to launch two versions.

The choice of technical route is critical. Apple is partnering with Nvidia to adopt NVLink Fusion interconnect technology, but its positioning is AI inference rather than training. Training requires the clustered computing power of thousands of GPUs, while inference places higher demands on energy efficiency. Apple's chip energy efficiency advantages have already been proven in phones and laptops, and now it wants to apply the same logic to data centers.

On the surface, it is a consumer electronics company returning to the enterprise market; underneath, AI computing demand is diverging. The training market is locked down by Nvidia, while the inference market has no absolute winner yet. Apple is not going head-to-head with Nvidia in training, but instead using energy efficiency as its entry point, attempting to tear open a gap on the inference side. If successful, enterprise users will buy Nvidia for training and Apple for inference, and the power structure of the computing supply chain will be rewritten.

(Sources: Bloomberg / Reuters / MacRumors / The Information)

4 | OpenAI Discusses Fundraising at $1.2 trillion Valuation, Just After CEO Said He Wanted to Slow Down

OpenAI is in discussions with investors about a new funding round at a valuation of approximately $1.2 trillion. The previous round in March valued the company at $852 billion, a gain of over 41% in half a year. According to reports, this round was initiated by investors, and OpenAI did not formally launch a fundraising process. Altman also explicitly ruled out the possibility of an IPO in 2026.

The timeline is noteworthy. Just two weeks ago, Altman published a long essay supporting AI deceleration, publicly agreeing with Amodei's call to "slow down the frontier," and announced that the IPO would be postponed to 2027. The deceleration statement and the trillion-dollar valuation funding discussions are proceeding almost in parallel.

Apparent contradiction, underlying coherence. The deceleration narrative reduces regulatory risk and public hostility while making competitors hesitate about whether to continue increasing their investments. For investors, a company that both calls for deceleration and can sustain a 41% valuation increase precisely proves its market monopoly power. The safety narrative and capital expansion are not opposites; in OpenAI's case, they are two sides of the same strategy.

(Sources: Bloomberg / PYMNTS / Quartz / Benzinga)

5 | SK Hynix in Talks with Intel to Produce Memory Chips in the U.S.

SK Hynix is in talks with Intel, considering using Intel's Ohio facility to produce HBM (high-bandwidth memory) chips. The two sides are also discussing bringing in a cloud computing company as a joint venture partner to share construction costs. However, the South Korean government may block the deal under the National Core Technology Act, which restricts the overseas transfer of core semiconductor technologies.

This is a significant expansion of the U.S. chip reshoring strategy. Over the past three years, from the CHIPS Act to TSMC's Arizona facility, reshoring has mainly focused on logic chips (processors). SK Hynix's talks mean that reshoring is extending from logic chips to memory chips. HBM is one of the key bottlenecks in AI training, and currently more than 90% of production capacity is concentrated in South Korea.

Intel's role in this potential deal is also worth noting. It is no longer the traditional positioning of chip designer and manufacturer, but is transitioning toward a "capacity lessor." Foundrying orders for TSMC, leasing facilities to SK Hynix, Intel is becoming an infrastructure company for the U.S. semiconductor supply chain.

(Sources: Quartz / TechCrunch / CNBC / Motley Fool)

Also worth knowing ↓

House tax committee advances crypto tax bill 38-5. After the Clarity Act stalled in the Senate, the House took a detour with the Digital Asset Tax Certainty Act to keep crypto legislation moving, with the tax path replacing the market structure path. (Source: CoinDesk / The Block)

Goldman Sachs research: U.S. labor income share falls to 52.8%, the lowest since 1947. BLS data shows that over the same period, the S&P 500 rose 600% while wages rose only 12.5%. The scissors gap between returns on capital and returns on labor may accelerate in the AI era. (Source: Fortune / BLS)

Jensen Huang will attend the state dinner Trump is hosting for Xi Jinping on September 24. In the same week he publicly opposed slowing down AI, Jensen Huang secured a seat at the state dinner. Chip export restrictions are one of the core issues at the U.S.-China summit. (Source: CNBC / Bloomberg)

Anthropic merges Claude chat and Cowork into a single AI assistant. It simultaneously launched two new products, Claude Docs and Claude Slides (beta), and fully opened the Design tool. AI companies are all moving from a single chat interface toward "super apps." (Source: Fortune / TechCrunch)

Novo Nordisk partners with Anthropic on AI drug discovery. A combination of the weight-loss drug giant and an AI lab. Anthropic's commercialization path is expanding from API services to deep vertical industry partnerships. (Source: Bloomberg / WSJ)

Bain Capital Ventures raises $1.6 billion, with the theme "After AGI." The venture arm of a long-established PE firm is betting an entire fund on the judgment that "artificial general intelligence has already arrived." (Source: Bloomberg)

House passes the Data Center Taxpayer Protection Act 417-3. Passed almost unanimously, it requires data center construction projects to disclose the use of taxpayer subsidies. Political resistance to AI infrastructure is escalating from the local level to the federal level. (Source: Axios)

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia