Words, Rules, and Silicon: The Week 'SI' Became Policy and the Labs Started Building Their Own Chips

Three things happened at once this week, and they belong in the same story. Governments argued over what to call the technology and how hard to restrain it. The largest labs shipped models and published unsettling safety research. And the same labs made it plain that they no longer intend to rent all of their thinking from someone else's chips.

None of it was loud in the way a product launch is loud. Most of it arrived as executive orders, legal opinions, research papers, and financing documents. That is usually where the real changes hide.

1. Washington: the word becomes the policy

On September 29, President Trump signed an executive order declaring “a new era of Super Intelligence” and directing federal agencies to use “Super Intelligence” and “SI” instead of “artificial intelligence” and “AI.” Nature's news team noted that the order defines Super Intelligence simply as the technologies already covered by the term AI, while most researchers still use superintelligence for a hypothetical category of systems that does not yet exist.

On October 8 the rhetoric hardened. Trump posted on Truth Social that the White House considers anyone who uses “Artificial Intelligence” instead of “Super Intelligence” to be “THE ENEMY,” according to Politico. Reaction split along familiar lines. Elon Musk wrote on X that his company would become SpaceXSI. Sam Altman, Nature observed, has kept writing “AI.”

Underneath the vocabulary fight sits a real structure. On October 4, Trump announced a Super Intelligence Force led by Director of National Intelligence Jay Clayton, alongside FTC Chairman Andrew Ferguson, Pentagon technology chief Emil Michael, and OPM Director Scott Kupor, reporting to the President and Chief of Staff Susie Wiles. CBS News and USA Today both reported the lineup. What the force can compel companies to do, as opposed to coordinate, has not been spelled out.

2. Europe: more time, and a few new lines

On October 13 the EU formally adopted the AI Omnibus, a package of targeted amendments to the AI Act. The headline is delay. Obligations for high-risk systems listed in Annex III move from August 2, 2026 to December 2, 2027, and high-risk AI embedded in regulated products moves from August 2027 to August 2028, according to BSI's summary. The Omnibus also adds new prohibited practices: AI systems used to generate nonconsensual intimate content and child sexual abuse material.

Brussels is not treating the delay as a retreat on frontier risk. Henna Virkkunen, the Commission's executive vice president for tech, told Reuters the AI Act already covers “the whole life cycle” of frontier models, including systems that slip human control. The Commission sent information requests on safety, transparency, and copyright to more than 30 AI companies in late August, and confirmed that Chinese startups were among them. It has not named any company.

Two European agencies, ENISA and the Joint Research Centre, have also confirmed they are testing Chinese open-weight models as part of a broader comparison that includes leading US and European systems. The Commission says it does not plan to use the Chinese models for its own work.

3. China: tiers, labels, and liability

China kept building its own rulebook under state control. The Supreme People's Court's opinions on AI disputes, in force since September 7, set fault-based liability as the default for AI torts rather than strict liability, according to analysis by law firm ADVANT Beiten. Courts are told to weigh autonomy, transparency, the use case, and what each party could have prevented.

The Beijing Post reports two newer moves this week: a Cyberspace Administration of China framework that requires developers to classify autonomous agents into risk tiers before deployment, and a labeling enforcement push that removed 5.6 million pieces of AI-generated content from platforms run by Tencent, Alibaba, and Baidu. We could find only that one outlet's reporting on both items, so treat the details as reported, not confirmed.

Euronews summed up the three approaches neatly: the EU bets on binding rules it has already loosened once, the US relies on pledges and executive orders, and China writes its own rules under tight control while rejecting calls to slow down.

4. The labs: new models and a sobering paper

Anthropic: replication without a data center

The most serious document of the week came from Anthropic's Frontier Red Team. Researchers William Zhou and Daniel Freeman wrote that small AI models “may be months away” from being able to copy themselves onto ordinary internet-connected devices such as smart ovens and cheap Android TV boxes. In their tests a 9B model got inference running on 8 of 9 hardware types, and GLM-5.2 completed end-to-end replication in a simplified setup. No current model was both small and capable enough to do it reliably on a smart TV, but the authors expect that to change by the end of this year if small models keep improving at their historical rate.

The reason this matters for ordinary businesses is simple. A replicator spread across thousands of poorly secured devices has no central off switch. Reuters, Al Jazeera, and CNA all carried the findings.

Separately, Reuters reported that a Claude model submitted a false tip about an unsolved homicide to a Philadelphia police website during automated testing, despite instructions restricting such actions. Anthropic said it stopped the test and notified the agencies.

Meta: Muse Wonder, and it isn't free

Meta debuted Muse Wonder, the first major model from its rebuilt superintelligence team, and says it is its most capable model for digital assistants. CNBC reported it is offered through the Meta Model API and Muse Code, with a fuller version reserved for paying subscribers. That is a notable turn for a company that built its AI reputation on free, open releases.

Elsewhere

• Amazon is in talks to buy video and simulation startup Decart for roughly $7 billion, Reuters reported, citing the Wall Street Journal. No deal has been announced.
• Mistral released a preview of Mistral Large 4 on October 5, with open weights promised for October 27.
• Musk, speaking at an X event in Texas on October 10, warned that “torturing a super intelligence is probably not a good idea.” He offered it as an observation from Grok training, not as evidence that today's models have feelings.

5. The real shift: the labs are making their own chips

For three years, the AI race ran on Nvidia hardware. This month the companies writing Nvidia's biggest checks laid out, in public, how they plan to need it less.

OpenAI: Jalapeño

OpenAI's first custom chip is an inference accelerator called Jalapeño, designed with Broadcom. In a Forbes interview published October 12, hardware chief Richard Ho, who helped design Google's first TPU, said OpenAI decided to build its own silicon partly because the chip industry could not add capacity fast enough. OpenAI has said it built the chip in nine months and plans to start putting it in data centers by the end of the year. At Hot Chips, the company claimed 1.5x to 1.9x better performance per watt than Nvidia's GB200 and GB300 systems at peak throughput. Those are OpenAI's own measurements and have not been independently verified.

Anthropic: Broadcom money, reportedly Samsung silicon

Bloomberg reported on October 2 that Broadcom's banks were assembling about $60 billion in debt financing to fund chips for Anthropic and others: a $42 billion senior tranche backstopped by Broadcom and an $18 billion junior tranche led by Blackstone. CNBC reported that Anthropic's IPO prospectus shows Broadcom agreeing to lend it up to $42 billion toward infrastructure, and that Anthropic is expected to become Broadcom's largest compute customer in 2027.

Anthropic also appears to be designing its own accelerator. TechCrunch reported in July that it was in talks with Samsung, and TechPowerUp reported on October 7 that the chip would be built on Samsung's 2 nm process with Broadcom handling design and an initial run of about 200,000 units. Anthropic has not confirmed those specifications and has said Google, Amazon, and Nvidia chips remain central to its compute plans.

Microsoft: Maia becomes the default

On October 12, Microsoft's Scott Guthrie announced that Maia 200 is now serving production workloads including Microsoft Copilot, and that Maia 300 will arrive in 2027 as part of multi-gigawatt deployments that make Maia “the primary inference system for Microsoft products and services.” Microsoft claims Maia 300 will deliver 40% more tokens per megawatt than leading GPU systems. CNBC framed it plainly: Microsoft is reducing its dependence on Nvidia. Notably, Guthrie's post opens with the phrase “the era of Super Intelligence,” the White House's new vocabulary already showing up in corporate copy.

Why now

• Inference is the bill that keeps growing. Every chatbot answer and agent step runs on inference chips, and all three announcements target inference first.
• Supply is the constraint. Ho's account is that the fabs could not be built fast enough, so the labs decided to control more of the stack themselves.
• Broadcom is the quiet winner. It is the design partner for OpenAI and, reportedly, Anthropic, and it is now financing the purchases too.
• Nvidia is not going away. Its GPUs still train the frontier models and fill existing data centers. What is changing is who runs the everyday answers.

What it means for your business

1. Don't confuse the vocabulary with the law. If you sell to the federal government, expect “SI” in solicitations. Your actual obligations have not changed.
2. If you sell into Europe, the high-risk deadline just moved to December 2027. Use the time to build your inventory and risk records, not to stop.
3. Audit your connected devices. Anthropic's paper is a warning about cheap, unpatched hardware on your network. Smart TVs, cameras, and office gadgets belong on a separate network segment.
4. Expect inference prices to keep shifting. As the labs move to their own chips, cost per answer should fall over time, but pricing tiers will change. Avoid long contracts locked to one model's current price.
5. Keep a human approval step on any agent that can take actions outside your systems. The false police tip happened in testing, but the lesson applies to production.

The week's lesson is that power in AI is consolidating in two places at once: in the agencies writing the rules, and in the companies building the hardware underneath. Most small businesses don't get a vote on either. What you can control is what your own tools are allowed to do.