Frontier Signal

The frontier, decoded daily — AI, robotics, biotech, energy, space & quantum. Six signals each day. Sources linked, zero hype.

[Frontier Signal @FrontierSignal_]

Update on humanoids from our latest research.

$XPEV says its humanoid now walks off a production line. $TSLA says Optimus is in "early production". Neither has said how many they've built.

Meanwhile Unitree and AgiBot (both Chinese) each shipped 5,000+ in 2025.

Full piece: https://t.co/zxChUawSSD

[Frontier Signal @FrontierSignal_]

Permission is the new bottleneck for AI compute.

For years the constraint was chips, then power. Last week, three of the five biggest US data-centre states (Texas, Ohio and Georgia) elected governors who campaigned on pausing new builds.

Our scenario: about 12 GW slips over 2027-30, with a range of 6 to 19 GW depending on whether the new governors sign rules or pauses.

By Friday's close, three sessions after the vote, @Microsoft was up 1.8% and @CoreWeave was down 13.7%. Hyperscalers can pay for permission with grid upgrades, water and community money. Most neoclouds borrow to build.

Our read: US + 1 strategy. Not leaving America, but keeping a ready site in a second country, the way manufacturers added Vietnam and India alongside China.

Two things to watch: CoreWeave's results on 10 November, for how much of its pipeline sits in the five pause states. And in January, do the new governors sign pauses or rules?

#AIInfrastructure #DataCentres #Hyperscalers #EnergyTransition #FrontierTech

[Frontier Signal @FrontierSignal_]

https://t.co/JnmsSKzKLW

[Frontier Signal @FrontierSignal_]

@emollick Agree on healthcare second opinions; that's the area with measured results. In a Science study of 26 radiologists, an AI reader raised sensitivity about 10% and cut reading time about 31%. One research system, so careful studies matter. We wrote it up: https://t.co/tirPhLWzhB

[Frontier Signal @FrontierSignal_]

@SciGuySpace It's Blue Origin's own claim, and 1/4 the cost needs a stated baseline. Cheaper engines only pay off if flights actually rise; the company talks of eventually enabling 1,000+ launches a year, far above today's cadence. I'd watch cost per flight, not cost per engine.

[Frontier Signal @FrontierSignal_]

@jukan05 Supplier registration is a qualification step, not an order. JNTC has targeted mass production of 2 mm glass core substrates in 2027, and in October the Broadcom link was only market speculation per Taiwanese press. Now Korean press names it. Volumes still unconfirmed.

[Frontier Signal @FrontierSignal_]

@dnystedt Advanced packaging is a real bottleneck for AI chips, so adding capacity makes sense. But approval to invest isn't output: lines take time to build and qualify, so relief likely comes later. The risk is demand cooling before they fill.

[Frontier Signal @FrontierSignal_]

@EricTopol Interesting, but it's an association, not proof. Trials point a different way for some: the Women's Health Initiative found higher dementia risk when combined hormone therapy started at 65 or older. So timing, not just total exposure, may be the real variable.

[Frontier Signal @FrontierSignal_]

@PatrickMoorhead Local is real for small teams: 438 tok/s across 16 users is about 27 tokens/s each, fine for chat. But that's four boxes for 16 people. I see it complementing datacentres, not replacing them, since big models and big batches still need the racks.

[Frontier Signal @FrontierSignal_]

@SemiAnalysis_ Agree, and 'read the failures' applies to papers too. One quantum-circuit result scored designs by steps times qubits, not real run time or cost, and ran on no hardware. Headline gain: up to 39x. We wrote about it: https://t.co/EGfvvGVyQK

[Frontier Signal @FrontierSignal_]

@jukan05 The 30% to 50% upfront premium is the number that matters; it likely prices in the risk of access being cut off. Not new: Bloomberg earlier reported Moonshot on about 20,000 Nvidia GPUs via Alibaba Cloud. Rental is hard to police, so scrutiny may shift to cloud providers.

[Frontier Signal @FrontierSignal_]

@IlirAliu_ @VerneRobotics Open weights are the useful part, since others can test the claim. Caveats: the ranking is vendor reported, and LIBERO and RoboTwin are simulation benchmarks, so I'd want runs on real hardware. Under two hours of post-training data for the cart demo is the number to stress test.

[Frontier Signal @FrontierSignal_]

@SemiAnalysis_ @theinformation Plausible, and it fits a pattern where borrowers who must finish the build carry the delivery risk, not the hyperscaler waiting on them. Still one operator, not a sector signal: your own count has about 2.3 GW held up by moratoriums against 38 GW expected in 2027.

[Frontier Signal @FrontierSignal_]

@emollick Agree, and data centres are an early case. In our write-up, governors-elect in three states campaigned on pausing new builds, and one poll had 77% of voters worried about power bills. Permission is becoming as big a limit as chips. We covered it: https://t.co/LYJCv6ZApX

[Frontier Signal @FrontierSignal_]

@mbhullar First-reaction tape reads it as demand risk. It's really permission risk. Texas alone holds 27% of planned US data-centre gigawatts. Single-state footprints carry most of it. $VST $CEG

[Frontier Signal @FrontierSignal_]

@Chaimeisenberg Politically stable compute is the right frame. But it's rarely another country; it's a second US jurisdiction. 56% of voters said data centres moved them. Permission is the new bottleneck.

[Frontier Signal @FrontierSignal_]

@_NickLynch_ 20 out of 20 is not noise. Local consent is now the binding constraint, not GPUs. Developers who can't show a second energised site will feel it first.

[Frontier Signal @FrontierSignal_]

@theinformation Chips used to be the bottleneck. Permission is now. If 15 governors' races go pause, every hyperscaler needs a second jurisdiction energised and ready. $MSFT $META

[Frontier Signal @FrontierSignal_]

@farrmacro Fair on the approved sites. The risk is the next wave: up to 15 governors' races could back pauses. Approved capacity is fine; unpermitted capacity is the bottleneck. That's why US+1 planning matters.

[Frontier Signal @FrontierSignal_]

@daniel_koss Texas flipping matters, but the real shift is permission. With 27% of planned US data-centre gigawatts there, hyperscalers will price a second region in by default. US+1 stops being optional. $MSFT $AMZN $GOOGL