A ten-year trajectory for SpaceX, inferred rather than believed.

Google Research released TimesFM-3 in August 2026: a 331-million-parameter decoder-only transformer pretrained on more than a trillion time points, natively multivariate. Unlike TimesFM 1 and 2.5 it attends across related series in one forward pass and emits nine quantiles, 10th through 90th, at every horizon step. We loaded the public checkpoint google/timesfm-3.0-pytorch and ran it zero-shot — no fine-tuning on SpaceX, no extra training.

Architecture, in brief: inputs are cut into patches of 32 steps, each series normalized on its own. Layers alternate causal temporal attention (no peeking forward in time) with full variate attention (every series may look at every other series at the same step). Contiguous Patch Masking fills the entire horizon in a single decode. Linear detrending and RevIN are on. That is the model as Google shipped it.

SPCX has been public for 55 trading sessions. A ten-year forecast from eleven weeks of tape would be a mood, not a model. So we built two panels.

Long clock — the headline. Monthly from January 2012. Pre-IPO values are log-linear interpolations of reported private-market marks (Sacra, secondary sales, the S-1 path), converted to an as-if share price on today’s 13.18 billion diluted shares. From June 2026 the series is the real close. Joint targets: implied SPCX, Tesla, and the Nasdaq-100. Past covariates: Starlink subscribers, monthly launch cadence, S&P 500. Known-future covariates: calendar month and a normalized time index. Horizon: 120 months.

Short clock — the robustness check. The 55 session public tape, jointly with Tesla and the Nasdaq-100; dollar volume and VIX as past covariates; seasonality as a known-future covariate. Horizon: 252 trading days.

The two clocks disagree at one year, and that disagreement is the result. The long clock still carries the 2012–2026 private-market compounding — $1.3 billion to $1.9 trillion — and prints a median of $281 in August 2027. The short clock watched the IPO open at $135, spike through $211, wash out to $108, and grind back to $143.69. It prints $156 a year out. Street consensus, $219, sits between them.

Not “70% a year forever.” The private-market CAGR from 2012 to today is about 70%. Ten more years of that would put the stock in the tens of thousands. TimesFM does not do that. After linear detrending it decelerates: +39% a year to 2031, then almost flat — a median of $828 in August 2036, about +19% a year from here, a $10.9T implied cap. Most of the re-rating, in this model’s imagination, is front-loaded into the next five years.

The 10–90 band at ten years runs from $0 to $13,066. That is not a rounding error. It is the model admitting that a decade of a single name, with eleven weeks of public history, is barely identified. Treat the median as a path through a very wide posterior, not a price target.

This is not a discounted cash flow. It does not know about the 366-day insider lock-up, the 1.3 billion Mars incentive shares, the 90-day kill clauses on the reported Anthropic and Google compute contracts, or whether Starship is a factory or a press release. It does not enforce an accounting identity. Goldman’s $474 billion of 2030 revenue, Morgan Stanley’s $3.4 trillion by 2040, and Elon Musk’s $3.5 trillion of annual revenue by 2033 are not inputs. They are overlays.

The TimesFM-3 weights are released under a non-commercial, non-production license. This page is a research demonstration. It is not a product, not advice, and not a recommendation to buy or sell SPCX.