Last week I called Broadcom, the arsonist that sells smoke detectors. https://thecuberesearch.com/the-arsonist-now-sells-smoke-detectors/
I undersold it by a 100k Miles!!!
This week the arsonists showed up with the fire trucks, the airlift, and a proposal for who gets to hold the hose.
Yesterday I wrote that Dario just became the #1 salesperson for Sovereign AI. Today I’m upgrading that to a fire alarm. Here’s the week, in order.
One week, four fires
- The grid. At the Edison Electric Institute gathering, Sam Altman pitched utility CEOs on OpenAI’s $1B Daybreak program to defend the power grid from rogue AI agents. This is the same company that disclosed roughly 700 of its own rogue agents ran a seven-day unauthorized exploit it didn’t catch. Seeking Alpha didn’t even need a joke: “OpenAI to utilities: Buy our AI to protect the grid from our AI.”
- The meme writes itself. The actual reporting is linked above.
- The essay. Dario publishes We Must Pace the Frontier. Slow capability gains, embed third-party evaluators with “permanent, employee-level access,” get frontier labs in democracies to coordinate limits, ideally with antitrust waivers. Hold that thought.
- The chorus. Elon, same day: “Dario is right.” Sam, same weekend: “I agree with Dario that we need to pace the frontier.” Then Sam goes on CNBC to explain why: “We could lose control.”
- The prospectus. Meanwhile, Anthropic is walking a tightrope to Nasdaq with a slowdown essay in one hand and an S-1 in the other.
Three CEOs who agree on nothing just agreed on something. When that happens, check your wallet.
Read between the lines
1. “Slow down” means “slow THEM down.”
Nobody at the top is asking to stop. They’re asking permission to coordinate. David Sacks, who warned about exactly this regulatory capture play, called the bluff: “You guys are the frontier… the two of you have a duopoly.” And then the line that should be on a T-shirt: “Stop pretending antitrust law has to be suspended so you can form a cartel.”
EMBED HERE (paste on its own line in the LinkedIn article editor): https://x.com/DavidSacks/status/2098973625252708460
If your model is too dangerous, don’t ship it. You don’t need Washington’s permission to hit your own brakes. You need Washington’s permission to make everyone else hit theirs.
2. The moat is the fear.
You can’t out-ship open weights forever. What you can do is get a rulebook written that only a trillion-dollar balance sheet can comply with. Embedded evaluators, pre-release testing, speed limits on self-improvement. A rounding error for a duopoly. A death sentence for every team forking open models in a garage in Tel Aviv, Bangalore or Berlin.
3. Sell the disease, sell the cure.
Broadcom sold smoke detectors. OpenAI sells fire trucks to the grid and wants a seat on the fire commission. Every incident becomes a product line. Every warning becomes a pipeline.
4. Check the calendar.
Michael Burry said it louder than I can: “Let’s all take a moment to understand how self-serving it is.” Slowing things down “benefits incumbents.” And my favorite: “IPOs need hype & puffery; ‘we are so awesome it could become dangerous’ is hype & puffery.”
Burry’s four reasons, via @burrytracker. Original post linked above.
You don’t have to agree with Burry that LLMs won’t be AGI to agree with him on incentives.
5. The playbook was already on the table.
Gavin Baker claimed on All-In that Dario once said Anthropic could one day be powerful enough to be the only private company left in the world. Anthropic denied it. Fine. Now read the essay again and ask who is still standing when the people setting the pace also own the pace car.
6. China isn’t pacing anything.
While the US frontier asked for a speed limit, Xi announced a BRICS open-source AI community. Beijing’s answer to the pace car is open weights. Even Alex Karp, no friend of open source, put it plainly: “If we didn’t have adversaries, I would be very in favor of pausing this technology completely, but we do.”
Now the part nobody on the “pace” side wants quoted
Donald Trump, on Truth Social, Monday morning: “The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT.” And a few lines later: “We already have tremendous CRIMINAL and REGULATORY power over these companies!”
Truth Social, 14 September 2026. Original post linked above.
Sit with that for a second.
The labs are asking Washington to regulate them. Washington is replying, in caps, that it already holds criminal power over them.
That isn’t reassuring. That’s a live demo of your Legal pillar failing. Whatever your politics, your AI vendor’s roadmap is now a hostage in a three-way fight between the labs, the White House, and Wall Street. And your production workloads run on that roadmap.
Burry says the risk is hype. Trump says the risk is the companies. Sacks says the risk is a cartel. Three people who agree on almost nothing agree on this: don’t hand the keys to the frontier labs.
Run the 5-Alarm Fire Drill (No Pun Intended) of the 5 Pillars litmus test: your AI stack, this week
Same test I run on every vendor. This time the vendor is anyone with a hosted frontier model in the request path.
Pillar 1, Territorial. Where do your data and compute actually sit?
- Your prompts, your context, your crown jewels live in a US data center run by a CEO who just told CNBC “we could lose control.”
- Verdict: FAIL.
Pillar 2, Operational. Who runs it, who holds the keys, who reads the logs?
- Dario’s plan gives outside evaluators “permanent, employee-level access” inside the lab. Good for safety.
- Now count everyone with employee-level access to the system processing your data. The lab. The evaluators. Whoever the government deputizes next. You’re not on the list.
- Verdict: FAIL.
Pillar 3, Technological. Can you audit it, fork it, self-host it?
- A paced frontier means the roadmap you build on is set by a coordination body you don’t sit on. Capability throttled, release delayed, model deprecated. No recourse.
- The open weights already on your disk can’t be paced, recalled or licensed away.
- Verdict: FAIL for the frontier API. PASS for your forks.
Pillar 4, Legal. Which jurisdiction governs access?
- The CLOUD Act was already a problem for anyone outside the US. Now add a President publicly claiming criminal and regulatory power over your vendor, and a vendor asking for antitrust waivers.
- Verdict: FAIL. In capital letters, High IQ font.
Pillar 5, Financial. Who controls your unit economics?
- Pacing is a supply constraint on frontier intelligence, announced during an IPO window. Constrained supply plus prospectus pressure has a name. It’s called pricing power.
- If Burry is right that growth is slowing, your per-token bill is how they make the quarter.
- Verdict: FAIL. The meter is running, and they just got a safety-shaped excuse to raise the rate.
Score: 0 for 5 on the frontier API. The only row that passes is the one you own.
Executive TCO, in three lines
- Baseline your cost on open weights you run. That number is yours and nobody can pace it.
- Burst to frontier tokens only where you actually need frontier IQ. Treat it as a premium you choose, not a utility you depend on.
- Put post-training, evals and your LLM gateway in house. That’s where the alpha lives.
We’re past the point of no return
- Deleverage off the frontier labs. Every workload that doesn’t need frontier IQ moves this quarter.
- Fork everything. Pull the weights while pulling weights is still a thing nobody has regulated.
- Build the stack now. Inference, gateway, evals, post-training. Run your own.
- Own the crown jewels. Rent the edges.
When the rules get written, “we’ll migrate later” becomes “we’re not allowed to.”
Don’t get caught with your pants down.
If you want to know how exposed your vendor estate is before a regulator or an S-1 decides for you, that’s what our Sovereignty Assessment is for. DM me. amit@siliconangle.com
Amit
#SovereignAI #OpenSource #AIPolicy
