The Adjacent Brief

TL;DR: SpaceX is fighting rival satellite operators over radio spectrum rights, a licensing contest DealBook frames as the real scarcity in orbit. Epoch AI projects the installed base of AI chips will double every nine months, the anchor number in a new Times interactive on the data center buildout. An AI-generated track charted on the Billboard Hot 100, and Gary Marcus argues OpenAI's Astra demo video was cut to flatter the model.

Worth Reading

• AI chips in use are on track to double every nine months — the buildout has no natural stopping point yet — Epoch AI's projection, mapped against the physical footprint that has to absorb it.
• Astra is a real advance and an edited demo at the same time — Marcus's read: the capability is genuine, the presentation is selective, and the gap between them is now a product decision.
• Google's AI opt-out doesn't stop your content from showing up in AI Overviews — Publishers' only real lever is full de-indexing, which is not a lever.
• Chinese VCs are raising again after three years at the bottom — the thesis is robotics and AI (paywall) — LP money returning to a market Western investors mostly wrote off.
• Digital detox has become a performance, and the people doing it know — In Bed With Social on why unplugging content is still content.
• Radio frequency sensing lets machines perceive a room without a camera in it — Bilawal Sidhu on the sensing layer that sidesteps the privacy fight cameras created.

Connected World

Heat is the ceiling — at 1mm and at 100 megawatts

Yanko Design reports that thermal dissipation, not battery capacity or optics, is what's holding smart glasses back — the proposed fix is a one-millimeter piezoelectric fan sitting against your temple. Scale that constraint up several orders of magnitude and you get the Times' interactive on the data center boom, where Epoch AI's projection that the number of AI chips in use doubles every nine months is really a projection about power draw and waste heat. The same engineering problem shows up in wearables as runtime and in Virginia as a substation queue. For anyone modeling AI cost curves: the line item that scales worst is cooling, and cooling is site-specific, permitted locally, and doesn't get cheaper with Moore's Law. It also explains why the physical-AI conversation keeps drifting toward edge deployment, where the workload is small enough to be air-cooled on the device.

In space, the asset is the license

Falling launch costs made orbital hardware cheap enough that the binding constraint moved upstream to the ITU and FCC filing cabinets, which is what DealBook's account of SpaceX versus rival operators over radio spectrum (paywall) is actually about. Spectrum rights go to whoever files and demonstrates use first, rewarding the operator that can put satellites up fastest — a self-reinforcing loop where deployment capacity converts directly into permanent regulatory position. Competitors are losing a paperwork race whose outcome is durable in a way that hardware advantage isn't.

The New Consumer

Moving home is a balance sheet decision, and it shows up in other people's revenue

More young college graduates are living with their parents (paywall), and the Times treats the question as cultural: is it a failure to launch, or a rational response to rent. The commercially useful angle is the demand mechanics: household formation is the trigger event for roughly a decade of first-time purchases, from leases and furniture to appliances, insurance, and renters' credit products. Deferred formation doesn't destroy that demand, it delays and compresses it, and it shifts the discretionary spend that would have gone to rent toward categories parents don't subsidize — travel, apparel, and going out. If your acquisition model assumes the 24-year-old is signing a lease, check whether it's still true.

The opt-out that doesn't opt you out

Search Engine Journal's walkthrough of Google's controls confirms the awkward part plainly: publishers who opt out of AI search features still have their content summarized in AI Overviews, because the only mechanism that truly excludes you also removes you from search. That's a choice between participating on Google's terms and disappearing, which is why publisher licensing negotiations have moved from traffic-sharing to compensation. A different flavor of the same design question shows up in Hank Green's admission that his own chatbot use is producing a dopamine loop he describes as unhealthy, notable less as data than as source. Green built a career reverse-engineering attention mechanics, and he's describing the engagement pattern from the inside. Self-report from one creator isn't a trend, but it rhymes with the digital-detox communities that have kept growing on Reddit and with Gen Z's drift back toward offline media, a thread we've been following for months.

Brand & Growth

Your comp plan pays for fires, so you get arsonists

The CS Café makes a sharp operational point: customer success review systems that reward dramatic saves systematically underpay the CSM whose accounts never break. Prevention produces no artifact — no escalation ticket, no war room, no executive thank-you email to screenshot in a performance review. The rep who lets an account decay to 40% health and then rescues it looks like a hero in the QBR deck; the rep who kept it at 90% all year looks like she had easy accounts. This is a measurement failure with a direct revenue consequence, and it generalizes well beyond CS: it's the same reason security teams get funded after breaches and infrastructure teams get funded after outages. If you own a retention number, the fix is available now: score leading indicators (usage depth, exec sponsor changes, ticket velocity) rather than outcomes, and pay for the flat line.

Commerce Rewired

Chinese capital is coming back, and the thesis is robots

After three years at record-low fundraising, Chinese VC firms are raising new funds against renewed appetite for domestic AI and robotics (paywall), per the FT's Eleanor Olcott. The number matters less than the thesis: LPs who spent three years treating China tech as uninvestable are underwriting hardware-adjacent AI, embodied systems, industrial automation, and components — not the consumer internet plays that defined the last cycle. That's a bet on the part of the stack China already dominates physically, and it sits alongside other signals we've been tracking, including China's positioning to capture the recovered critical minerals coming out of end-of-life EV batteries. For Western operators, the practical implication is on the cost curve for robotics BOM over the next 24 months, not on any near-term competitive threat in software.

Culture & Signal

The charts are becoming a certification body, and the labels are writing the rules

The Verge's investigation into whether a Fenix Flexin track that reached the Hot 100 is AI-generated is more interesting for what it can't resolve than for what it finds: there is no forensic test that settles the question, only vibes, metadata, and denials. That vacuum is precisely why the three major labels have started floating chart-eligibility rules that would exclude AI-generated music. Read the incentive: the majors don't have a detection method either, but they control the definitional machinery, and a rule that requires provenance documentation advantages the party with contracts, session files, and lawyers. Human-made becomes a credential that costs money to prove, a moat dressed as an ethics position. Anyone building a provenance or attestation product should be watching who ends up administering the standard, because that's where the pricing power lands.

Machines & Minds

Checkable results beat cinematic ones

Gary Marcus's dissection of OpenAI's Astra launch grants that the model is a real advance and then documents how the demo video uses cuts and framing to imply capabilities the model doesn't reliably have. Set that next to Tyler Cowen's roundup of fresh mathematical results produced with GPT models, where the claims survive because someone else can verify the proof. Same underlying systems, opposite epistemic footing: one is evaluated by whether it looked impressive, the other by whether it's correct. The useful discipline for buyers is to insist on the second condition: in procurement, that means no capability claim enters the evaluation unless your own team can run the check on your own data. Demo-driven diligence is how organizations end up with the deployment gap Adobe hit when it discovered no vendor could actually adapt its B2B sales playbook and it had to build the thing internally. The capability is often real. The packaging is doing work the capability can't.

11 articles across 6 themes · 10 sources · Powered by Folo + Claude