AI social digest — Oct 25, 2026
Top line: AI economics dominated the Musk channel — and now carry its founder's balance sheet: in a reply on October 25 (~3:11 PM ET), Musk said debt markets have "almost entirely shut off access to money" for companies without AI-independent revenue, stated "SpaceX closed $40B of debt for GPUs," and predicted "guppy-level death and consolidation" for the 200-plus neoclouds, capping a weekend of amplified a16z GPU-rental charts, David Sacks's per-gigawatt SpaceX-vs-CoreWeave math ($50 billion with a one-year payback vs $20 billion at 2.5) and a "stranded compute" warning. The day's other currents are agent-shaped plus one measurable: Grok Bot's announced phone calling (v0.77) picked up a third-party validator, Perplexity's SF 2049 roadster game went multiplayer, OpenRouter put Anthropic-model token volume up 3.6x since Opus 5.5 launched, and in a Fireside Alpha conversation Musk sketched AI satellites built from lunar silicon and aluminum — while a $1 X Money long-press tip was announced with no rollout detail.
- GPUs age like fine wine — until someone asks who keeps paying
- Sacks's per-gigawatt math: SpaceX at $50B and a one-year payback vs CoreWeave's $20B and 2.5
- Grok Bot announces phone calling: hold music, key-presses, a live transcript
- SF 2049 goes multiplayer: race rooms, agent-made cars, an agent-shot trailer
- Fireside Alpha: AI satellites fed by lunar silicon and aluminum
- Musk's debt-market read: SpaceX "closed $40B of debt for GPUs," the neoclouds face "guppy-level death"
- OpenRouter's third-party usage read: Anthropic token volume up 3.6x since Opus 5.5 launched
GPUs age like fine wine — until someone asks who keeps paying
Sources Elon Musk (repost of NVIDIA AI Infrastructure) Elon Musk (stranded compute) Elon Musk (depreciation reply)
What happened: Musk amplified two a16z chart posts on X on October 24–25, 2026: rental rates for NVIDIA GPUs are up across the board since January 1 — a16z's approximate year-to-date figures are H200 +160%, B200 +120%, RTX PRO 6000 +65%, H100 +35%, with even the six-year-old A100 renting above its price at the start of the year — and, zoomed out, GPUs still renting ahead of a standard six-year depreciation schedule (H200 +143%, B200 +111%, RTX PRO 6000 +73%, H100 +54%, A100 +9%, with most of the gap opening after April as token demand took off). Musk's caption: NVIDIA AI infrastructure "is a durable, investable asset that retains value." Hours later he volunteered the downside in his own voice, replying to a question about data-center supply: "I think we're going to see a lot of stranded compute in partially complete data centers." And on the reply chains the next morning (October 25, ~11:43 AM ET), Musk pushed back on the depreciation framing itself in the debate the charts kicked off: "Accounting depreciation is irrelevant (true depreciation is not). Delivering on projected cash flow is what matters" (Elon Musk (depreciation reply)).
What's genuinely new: Specific dated GPU-rental percentages, including a six-year-old chip renting above its January 2026 price and rentals beating the standard depreciation schedule — concrete numbers on the storyline's "who pays for the buildout" question, arriving the same weekend SpaceX's own disclosure that AI infrastructure brought in $7.6B (52% of Q3 revenue) was already circulating. Musk's stranded-compute line is the first oversupply warning from the loudest name in the buildout itself. All of it is X-grade commentary and chart-craft, not measured market data; the October 22 Axios token-price charts remain the only empirical pricing read so far.
Watch: Whether a GPU-rental index joins token prices as a quoted benchmark, whether the stranded-compute worry shows up in any of the financing storylines (CoreWeave-style distressed compute debt, Terafab-style buildout plans), whether the a16z rental figures get corroborated by a second source with methodology, and whether the cash-flow-over-depreciation argument reaches any mainstream analyst treatment.
Sacks's per-gigawatt math: SpaceX at $50B and a one-year payback vs CoreWeave's $20B and 2.5
Sources Elon Musk (relay of David Sacks's All-In readout)
What happened: On the evening of October 24, 2026, Eastern (relay published shortly after 10:10 PM ET), Musk amplified a post relaying David Sacks's figures from the All-In podcast: SpaceX takes in roughly $50 billion in monetization per gigawatt of AI compute with about a one-year payback, against CoreWeave at roughly $20 billion per gigawatt with about a 2.5-year payback. The spread is attributed to contract structure — CoreWeave leans on long-term customer commitments to finance its infrastructure, while SpaceX's balance sheet lets it run shorter contracts and capture today's elevated compute pricing. The same post cites a SpaceX target of roughly 13 GW of operational compute capacity by the end of 2027. Musk's overlay: "This is a fuller list of the limitations of supercomputer buildout. If anything is missing, however seemingly trivial, the compute doesn't work."
What's genuinely new: The first per-gigawatt monetization-and-payback comparison between SpaceX and CoreWeave — the exact financing contrast behind CoreWeave's slide into distressed bond territory on October 19 and Sunday's stranded-compute warning — plus a dated ~13 GW capacity target. Podcast-grade commentary relayed on X, not company filings, and the third Musk-channel economics datapoint in as many days after the $7.6B/52%-of-Q3 AI-infrastructure disclosure and the a16z rental charts.
Watch: whether the ~13 GW target and the per-gigawatt figures get a company-statement origin, whether any of it meets the stranded-compute question in financing coverage, and CoreWeave's next bond pricing or refinancing.
Grok Bot announces phone calling: hold music, key-presses, a live transcript
Sources Elon Musk (announcement) Elon Musk (relay of Jensen Huang) Elon Musk (RT of Brian Roemmele)
What happened: On X on October 25, 2026 (~10:35 AM ET), Musk announced (Elon Musk): "Grok @Bot can now call someone or a business for you, make a request on your behalf, take notes on the response and send you a note summarizing the conversation!" The same post carries an itemized relay naming version 0.77: a live transcript "while you watch" as the bot dials someone, sits on hold, presses keys and gets transferred; a Messages-on-Mac setup screen under Connect Apps that is described as "still switched off"; and Tailscale — the bot joining the user's private tailnet — named alongside them. A day earlier (October 24, ~11:06 AM ET), Musk relayed NVIDIA CEO Jensen Huang: "GrokBot is super cool. The demand for GrokBot, the consumption for GrokBot is really skyrocketing" (Elon Musk (relay of Jensen Huang)). An afternoon RT (October 25, ~1:49 PM ET) added a validator's voice: Brian Roemmele says he and "dozens of clients" already use the calling daily "with an external MPC," calls "this update to Grok Bot with integrated calling" spectacular, and promises "a whole slew of Grok Bots I will release that will make calls" (Elon Musk (RT of Brian Roemmele)).
What's genuinely new: Phone calls as an announced capability — the bot acting on a user's behalf over the public telephone network with hold behavior, input relaying and a post-call note — plus two further surfaces (Mac Messages, Tailscale) named as not yet switched on, and the version-0.77 numbering now echoed by a third-party power user whose own daily setup runs through an external middle process. Roemmele's account is advocacy-grade testimony, not documentation. Announced only: no changelog, rollout detail, or demo in the reporting.
Watch: Whether calls ship as described, whether Mac Messages and Tailscale go live, the changelog trail for v0.77, and the OS-permission wall (Apple's October 2 full-disk clamp is the standing precedent) if phone-side autonomy widens.
SF 2049 goes multiplayer: race rooms, agent-made cars, an agent-shot trailer
Sources Aravind Srinivas (X post)
What happened: On October 25, 2026 (~11:02 AM ET), Perplexity CEO Aravind Srinivas announced that SF 2049 — the Cyberpunk-inspired Roadster browser game at 2049.pplx.app first shown October 18 — "is bigger now, and it's multiplayer" (Aravind Srinivas (X post)): race rooms running on Cloudflare Durable Objects ("create a room, share a six-letter code," server-side lap timing, rooms closing when empty); ten new player cars generated with image-to-3D plus a studio garage; a Blade Runner-style flying Spinner; Rapier vehicle physics with handbrake powerslides; and a fullscreen map rebaked to the real San Francisco street layout. The trailer, the post says, was "shot by three agents in parallel" — fixed-step 1080p60 capture, a real two-player race on a second GPU, and sound rebuilt from the game's own engine loops — with "everything in it is captured gameplay."
What's genuinely new: The first update to the game since its October 18 debut — multiplayer race rooms on server authority, agent-generated vehicle assets, and a trailer claimed to be produced end-to-end by three parallel agents. Announced by the company's own CEO with a live URL, which puts it in a stronger evidentiary position than most announcement-grade claims of the weekend, but the parallel-agents trailer mechanics are still only asserted.
Watch: Verification of the three-agent trailer pipeline, server-side timing under real multiplayer load, and whether SF 2049 remains the game's living home as a Perplexity-side echo of the flying-Roadster story.
Fireside Alpha: AI satellites fed by lunar silicon and aluminum
Sources Elon Musk (Fireside Alpha relay)
What happened: In a Fireside Alpha conversation relayed on X at ~12:27 AM ET on October 25, 2026, Musk explained building AI satellites from lunar material (Elon Musk (Fireside Alpha relay)): lunar soil is "20% silicon or something like that," so silicon can be mined, refined and turned into solar cells on the Moon, with radiators made from lunar aluminum; the chips are "pretty light" to send from Earth — "but maybe at some point you make them on the moon, too." His scaling rationale: "I don't see any way that you could do 500 to 1,000 terawatts per year launched from Earth" — a "video game situation where it's difficult but not impossible to get to the next level." The same relay includes NASA's Artemis launch lead Jeremy Parsons calling SpaceX's between-flight Starship rebuild pace "unthinkable" in standard industry history.
What's genuinely new: The lunar-manufacturing layer of the space-data-center compute plan made explicit — cells and radiators from lunar material, chips from Earth at first — plus a dated 500–1,000 TW/year Earth-launch ceiling framing and a named NASA figure's attestation on Starship turnaround. Speech, not hardware or filed plans.
Watch: Whether the lunar solar-cell pitch reaches any SpaceX program or procurement document, any corroborating coverage of the Parsons attestation, and how the space-compute lane's scaling math meets Earth-launch or Terafab realities in later coverage.
Musk's debt-market read: SpaceX "closed $40B of debt for GPUs," the neoclouds face "guppy-level death"
Sources Elon Musk (debt-market reply) Elon Musk ("not buying any neoclouds")
What happened: In a reply on X at ~3:11 PM ET on October 25, 2026, Musk wrote (Elon Musk (debt-market reply)): "As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn't have substantial revenue independent of AI data centers. SpaceX closed $40B of debt for GPUs, but we're investment grade and have Starlink & launch revenue. That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many." In a separate reply the same evening (~5:01 PM ET), he ruled out buying capacity instead of building: "SpaceX is not buying any neoclouds. Pointless" (Elon Musk ("not buying any neoclouds")).
What's genuinely new: The first first-person statement that the $40B in debt for NVIDIA chip purchases first reported as sought on October 7 has closed, plus an investment-grade self-characterization resting on Starlink and launch revenue — and an outright prediction that most of the 200-plus neoclouds fail or consolidate, sharpening the same weekend's stranded-compute warning into a consolidation call. The "closed," "investment grade," and neocloud-count figures all rest on this single post; no filing or second source is in the reporting.
Watch: Any filing or independent reporting confirming the $40B closing and its terms, whether "guppy-level death" language reaches CoreWeave and neocloud financing coverage, how the consolidation prediction sits against the October 19 CoreWeave bond-distress read, and whether the refusal to buy neoclouds draws response from the companies themselves.
OpenRouter's third-party usage read: Anthropic token volume up 3.6x since Opus 5.5 launched
Sources nichochar (X post)
What happened: On X at ~7:17 PM ET on October 25, 2026, nichochar posted (nichochar (X post)): "can confirm that in developer circles anthropic is making a huge comeback, fueled largely by opus 5.5 — it orchestrates subagents with the other two which is why they also are growing." In the same exchange, Peter Walker relayed a new OpenRouter figure: the 7-day average token volume for Anthropic models is up 3.6x since the launch of Opus 5.5, with Sonnet 5.5 and Haiku 5.5 "also well received."
What's genuinely new: The first quantified third-party demand datapoint for Anthropic's completed 5.5 line (Opus 5.5 September 22, Sonnet 5.5 September 28, Haiku 5.5 October 7) — a 3.6x jump in weekly OpenRouter token volume since the Opus 5.5 launch — against adoption notes that had run vendor-adjacent, such as Perplexity Computer's day-one tier placement. OpenRouter is a neutral distribution channel but only a slice of overall usage, and the "huge comeback" line is commentary, not measurement.
Watch: Whether the OpenRouter figure gets cited in mainstream coverage as a usage benchmark alongside the GPU-rental and token-price reads, whether Anthropic confirms demand in its own words, and whether the "orchestrates subagents" pattern — Opus 5.5 steering Sonnet and Haiku — shows up in how developers deploy the line.
Brief
- X Money tipping on a long-press — Musk announced on X that a long-press on the like button will send a person $1 via 𝕏 Money (posted early October 25, Eastern; Musk's post). A small payments feature on the platform, but it continues the payment-rails march already under way — X Money entered the Grok Bot orbit with the October 21 finance-report integration, and the bot side already carries X Money/Stripe Link payment handling. Announced only; no changelog or rollout detail in the reporting. Watch: whether it ships as described and whether agents can trigger it.