Support Us & Upgrade to the Operator Edition

The AI trade developed a split personality this week. A Korean data centre company lost a third of its value on day one, while SpaceX told Wall Street its AI unit is now the main event. Same industry, same week, opposite verdicts, and the gap between them tells you exactly where the money thinks this is heading.

How do you maintain your distinct brand qualities while answering the endless velocity, volume, and variety demands?

Join our honest discussion on this evolution with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready-to-run prompts.

Join the talk

For years, AI video and AI voice lived in separate tools. You'd generate a clip in one, the voiceover in another, then fight with lip-sync. ElevenLabs just stitched them back together.

Chimera pairs Eleven v4 (their most emotive voice model yet) with Seedance 2.5 for video, in one flow inside ElevenCreative. Character in, talking video out.

Here's how:

🔍 Where this pays off:

💡 Pro tip: Direct the performance, don't just write the script. v4 follows audio tags and direction prompts more closely than earlier models, so add cues like [laughs] or [whispers] where the emotion should shift. A flat script gets a flat character.

Want the version of this that actually runs in your business?

The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.

This week: How to save money on lawyers and document SOP’s with AI.

👉 Upgrade to The Operator Brief

🚗 GM wants to cut vehicle development from 4-5 years to roughly two years using AI and simulation, and it's bringing eyes-off highway driving to the 2028 Escalade IQ. In a couple of years, you’ll be able to take your eyes off the road and keep them peeled on what GM will be cooking up next.

🧮 A new argument says that in the age of AI, learning math matters more, not less, putting it at odds with Jensen Huang's recent suggestion that kids may not need basic math. The math behind AI and machine learning, like processing data, learning from patterns, and making decisions, is here to stay—whichever way you calculate it.

⚖️ Anthropic pitched lawmakers a safety bill that would let the government block a model only if it risks 50+ deaths or $1B in damage, and critics are already doing the per-life math. This isn’t exactly the math I meant when I described the math behind AI above. This math is a bit messier.

📉 Korean AI data center firm Elice plunged over 30% on its stock market debut. Investors have been writing blank checks for AI infrastructure all year. Elice is a sign the blank checks are getting a second look.

🧬 OpenAI acquired Orchestra Bio to put AI agents inside drug development, pairing it with its Rosalind life sciences tools. The distance between “write me an email” and “help discover a medicine” has closed faster than anyone predicted.

Several Google DeepMind staff have resigned, or are on their way out, over the company’s growing military AI contracts, according to a BBC Eye investigation.

The trigger is a deal Google signed with the Pentagon earlier this year, which opened its AI models up to classified military work.

Liguori spoke out despite a contract that she says bars her from criticizing Google for the rest of her life. Staff told the BBC that people inside are afraid of retaliation. Google says it has channels for raising concerns and strict anti-retaliation policies.

The backdrop matters here. Defense Secretary Pete Hegseth wants the US to become an “AI-first” fighting force, and the Pentagon’s main targeting platform, Palantir’s Maven Smart System, already plugs in models from several leading AI labs.

Here’s the uncomfortable part:

Everything hinges on the phrase “appropriate human oversight.” Nobody outside the Pentagon gets to define what appropriate means.

Sarah Shoker, who used to lead OpenAI’s Geopolitics team, put the problem bluntly in the BBC piece: when AI selects hundreds of targets and the clock is short, a human can’t meaningfully check the work. The former head of Project Maven, retired Lt. Gen. Jack Shanahan, said the current philosophy seems to be speed first, testing second.

So a human is technically in the loop. Whether that human is actually reviewing anything is another question entirely.

🔮 Prediction: These resignations won’t change Google’s contracts. Not one bit.

Here’s why. Every major lab except a few has signed on, the government is actively pushing for it, and defense is one of the few AI customers with essentially unlimited budget. Walking away would mean handing that revenue and influence to OpenAI, Microsoft or SpaceX. No board is approving that because a handful of researchers left.

What it will change is who works where. One current DeepMind employee told the BBC this could be a “changing of the guard”: mission-driven veterans leave, early-career hires fill the seats. Over two or three years, that quietly rewires the culture of the lab building some of the most capable models on earth. That’s the real story, and it’s slow enough that nobody will notice until it’s done.

Two things I’d watch:

1. Gag clauses become the next fight. OpenAI got burned in 2024 over restrictive exit agreements and walked them back. A lifetime ban on criticizing your employer, applied to people who know how military AI is actually used, is going to draw attention from lawmakers in the US and the UK.

2. “Human oversight” gets defined by an incident, not a policy. Notice what’s missing from Anthropic’s proposed safety bill: it covers bioweapons, cyberattacks, and models escaping control, but not battlefield use. Right now, nobody is writing that rule. My bet is the definition arrives the hard way, after a publicly reported targeting error traced back to an AI system.

💬 If you worked at Google DeepMind, would you quit? Reply YES or NO. I read every reply, and I’ll share where you all landed next week.

Read the full BBC investigation here

SpaceX blasted off past its Q3 expectations, and rockets weren’t the reason.

Its AI division, now renamed Super Intelligence, grew more than ninefold and now makes up over half of group sales, according to Hargreaves Lansdown’s breakdown.

Shares jumped 8.1% after hours.

The business model is simple: SpaceX builds data centres and rents out compute on short-term contracts. It expects to end the year with roughly 3 GW of capacity and add another 10 GW in 2027, with rental rates still climbing. HL’s analysts think 2027 revenue estimates need to at least double.

Now hold that next to the Elice IPO flop. Same product (rented AI compute), totally different reception. One analyst summed up the mood shift: investors stopped asking “how fast can you build” and started asking “who’s paying, and for how long.”

Here’s the uncomfortable part:

SpaceX’s free cash outflow hit $10B in a single quarter, on $19.6B of capex. Q4 capex is guided at $35-40B.

HL estimates each gigawatt costs around $30-50B to stand up. Do the math on 10 GW in 2027, and you’re looking at $300-500B. HL thinks SpaceX will need to raise more cash within six months.

And SpaceX isn’t the only AI giant showing scary numbers this week. Anthropic’s draft IPO prospectus revealed an $80.7B net loss for the first half of the year. Before you panic, about $72B of that is a paper loss from revaluing convertible notes, which goes up as the company’s value goes up. The number that actually matters is buried underneath: Anthropic spent nearly $9B on compute in six months, up from $2.3B a year earlier, while revenue hit $16.3B.

Different companies, same story. The headline losses are noise. The compute bill is the signal, and it’s climbing faster than almost anyone planned for.

Money isn’t even the hardest part. Land, power, and chips at that scale are physical bottlenecks, not spreadsheet problems. And SpaceX investors get almost no say in how it plays out: Musk holds around 85% of the voting rights.

🔮 Prediction: The short-term rental model is both SpaceX’s genius and its biggest trap.

Right now, compute is scarce, so short contracts let SpaceX reprice upward every few months. That’s why margins look incredible this quarter. But short contracts cut both ways. The moment supply catches up, those same contracts reprice downward just as fast, and there’s no long-term commitment to cushion the fall.

That’s exactly what the market is punishing in neoclouds like Elice and CoreWeave this week: heavy debt, short customer commitments, and no structural edge once the scarcity premium disappears.

Here’s the twist: Anthropic’s spending is SpaceX’s revenue. Every AI lab burning billions on compute is a customer for someone like SpaceX. As long as labs like Anthropic keep growing revenue fast enough to justify that spend, the landlords win. The day one of them blinks and cuts its compute budget, every short-term contract in the market gets repriced at once.

SpaceX’s real defence isn’t the contracts. It’s cost and speed. It can build faster than almost anyone, Starlink throws off cash to fund the buildout, and Starship (if the Flight 15 dual catch lands next month) opens the door to orbital compute, a story no other cloud provider can tell.

So here’s what I expect: a major capital raise before mid-2027, and a quiet shift toward longer customer contracts at a discount. Watch for that second one closely. When the most aggressive AI landlord on earth starts trading price for duration, that’s your signal the compute scarcity window is closing.

The winners in the next 18 months won’t be whoever builds the most. It’ll be whoever can bring power online cheapest when the music slows down.

💬 Is SpaceX an AI company or a space company now? Hit reply with one word.

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI