Cross-cutting rail · the credibility engine
The receipts.
Dated, sourced predictions, tracked to resolution. Anyone can have a take; this is the take, graded over time. The ledger is the memory — the whole thesis, demonstrated instead of asserted.
Pending
Cursor reaches ~$6B ARR by end of 2026
SpaceX–Cursor $60B all-stock acquisition closes
Orbital compute for training and inference will be gated by cross-satellite scale-out bandwidth, because a satellite is approximately one rack of GPUs, and models are larger than one rack — so they must cross an inter-satellite link that is worse than terrestrial cross-rack networking.
Primary hypothesis. Four falsifiers tracked below. First hard data points: Suncatcher 2027 prototype, Starcloud-2 Jan 2027. Rubin-gen feasibility (2026-06-16): Rubin NVL144 rack ~600 kW (~4.5× GB200); solar ~1,500 m² (~6× Starlink V3); radiator 500–857 m²; compute-sat ~10.5 t (~9.5 per Starship, mass-limited). One satellite = one NVL576 scale-up domain (576 GPUs / 365 TB HBM4e). Frontier training spans multiple domains → orbital training remains constrained by the ISL cliff; inference-only serving is viable.
Falsifier 3: If asynchronous or low-communication training matures, the synchronous-step-gated-by-slowest-tier argument loosens for orbital training specifically.
Track low-comms training research (async SGD, local SGD, DiLoCo variants).
Falsifier 4: If active-parameter fractions keep falling (below ~3%), ~3-25 GB/s inter-satellite links become tolerable for inference-serving — confirming inference-first read while softening training-constraint claim.
Track sparsity_ratio in model_releases; orbital inference already plausible for models fitting one satellite domain.
Falsifier 2: If power/thermal advances plus larger NVLink-class domains let a frontier model fit inside one satellite, cross-satellite scale-out stops being the binding constraint for training.
Track per-satellite scale-up domain size; Rubin Ultra NVL576 as the upper bound for single-satellite packaging.
Falsifier 1: If deployed inter-satellite laser bandwidth climbs toward the 1.6 Tbps DWDM bench figure in production hardware (vs ~25-200 Gbps today), the orbital scale-out cliff shrinks and the training-constraint hypothesis weakens.
Watch Suncatcher DWDM deployment; Starcloud-2 laser specs at launch.
Cursor closes xAI's coding gap with frontier labs via in-house data + captive compute
Grok 4.3 total parameter count is disputed: described as both ~500B and ~1.5T in corpus. Neither figure is verified by xAI.
Do not use a single parameter count for Grok 4.3 until xAI discloses. Track 500B vs 1.5T conflict. params_disclosed=false in model_releases.
All-stock structure ties Cursor's realized value to post-IPO SpaceX share volatility
SpaceX scales Grok to multiple trillions of parameters (Grok 5, ~6T rumored)
SpaceX compute-leasing run-rate ~$26B/yr (Anthropic + Google) sustained through 2029
Anthropic completes an IPO at a valuation near or above its ~$965B confidential filing mark
Disaggregation of prefill/decode extends GPU useful lives to ~10-15 yrs and lowers financing toward 5-6%
Frontier labs eventually stop releasing top models via API (new prisoner's dilemma)
Nvidia could build a near-frontier in-house model whenever it wants (though likely won't)
Google stays well-positioned (most compute, YouTube/robotics data, search); Google I/O is a key Pareto-frontier test
Amazon shows real P&L efficiencies from robotics in retail over the next ~18 months
Near-term power/watts shortage begins to ease as new energy sources come online
Orbital compute (racks in space linked by lasers) becomes real and solves power long-term, mainly for inference
TSMC's capacity decisions are the single most important variable for whether AI gets a bubble
TSMC stays in a 'Goldilocks' zone — expands enough to block a >30%-share second source, not enough to overbuild
Intel or Samsung eventually breaks capacity discipline, forcing others to follow
If TSMC fully met Jensen's demand, Nvidia could sell ~$2-3T of GPUs in 2026-27 (but would trigger overbuild)
Terafab (SpaceX/Tesla + Intel) succeeds in building the world's largest fab in America
OpenAI and Anthropic each exceed ~$200B ARR this year via usage-based pricing + more compute
If Anthropic were unconstrained on compute it would already be doing ~$100-200B vs ~$50B ARR
Frontier tokens keep capturing the majority of model-layer economic value
A 'bitter-lesson' violation via ASI self-optimization is the single biggest risk to the AI trade
Continual learning arrives soon, producing a fast takeoff
Anthropic & OpenAI each reach ~5-6 GW
ASML/EUV becomes the #1 AI-compute scaling constraint
Elon Musk: build a 'TeraFab' — 1M wafers/month
Google must double AI serving capacity every 6 months
Elon Musk wants 100 GW/yr of compute in space
Sam Altman wants 52 GW/yr (~1 GW/week)
Space data centers don't make economic sense this decade
Memory prices double or triple again
Smartphone volumes fall to ~800M then 500-600M
China fully indigenizes DUV; EUV 'working' but not volume
~700 EUV tools installed -> ~200 GW/yr AI-chip ceiling
~half of new power capacity will be behind-the-meter
Anthropic & OpenAI each reach ~10 GW
xAI targets ~1 million GPUs at Colossus by late 2026