Revenue of $7.81B, Capex of $18.4B — In the Same Quarter

The headline results (all figures on a Q2 2026 basis, from the company's report and the August 4 earnings call):

Stretch the timeline out and this is the steepest stretch of the curve in two years. Full-year 2025 revenue was $18.674 billion with an operating loss of $2.589 billion and adjusted EBITDA of $6.584 billion; first-quarter 2026 revenue was $4.694 billion with an operating loss of $1.943 billion and adjusted EBITDA of just $1.127 billion (both from the prospectus filed May 20). Then the second quarter jumped to $7.81 billion of revenue and $3.5 billion of adjusted EBITDA. What actually bent the curve was the AI segment — first-quarter revenue of $818 million against an adjusted EBITDA loss of $609 million, second-quarter revenue of $2.56 billion against positive adjusted EBITDA of $1.1 billion. That is 213% sequential revenue growth and a swing from a $609 million loss to a $1.1 billion profit in three months. When a business flips like that in a single quarter, there is usually only one explanation available: the incremental revenue carries almost no incremental cash cost, because the cost has been capitalized onto the balance sheet. We will do that arithmetic later.

The connectivity numbers need a finer breakdown. Starlink ended the quarter with 12 million subscribers, double the 6 million of a year ago, but monthly revenue per subscriber fell from $85 to $66, down 22% year over year and barely flat sequentially (Q2 earnings call). Twelve million subscribers times $66 times three months is roughly $2.4 billion, while the segment reported $4.29 billion in total — the $1.8 billion difference comes from enterprise and government customers, up 108% year over year. The consumer side is trading price for international subscriber count; the enterprise side is backfilling the price the consumer side gave up. That is the real structure of this segment's growth right now.

So which number matters in the next report? Consensus full-year revenue currently sits at $44.63 billion (35 analysts, per stockanalysis), while the first half actually delivered $12.50 billion ($4.694 billion plus $7.81 billion). That leaves $32.1 billion to be done in the second half, an average of $16.0 billion per quarter — more than double what the second quarter produced. Management's framing is a $100 billion annualized revenue run rate by the end of December (CFO Bret Johnsen, Q2 earnings call), which works out to $8.3 billion in the month of December alone. Put those two together and the third quarter has a hard pass mark: roughly $12.0 billion of revenue, a 54% sequential increase. Come in below that and the $44.63 billion full-year consensus has to be marked down — and the current valuation is priced off exactly that consensus.

The demand-side evidence, for now, is solid enough. Of the two large compute tenants, Anthropic has taken roughly 325,000 Nvidia GPUs at $1.25 billion per month and Google roughly 110,000 at $920 million per month, together about $26 billion annualized (Fortune 2026-07-19); Department of Defense data center capacity is still under discussion and unsigned. Musk said on the call that the binding constraint on compute has moved from chips to memory, with demand growing 200% a year against supply growing 20%. Gwynne Shotwell said the company has not lost a single enterprise customer. Demand is not where the problem is.

That $18.4 billion from the last chapter deserves a direct answer. Second-quarter adjusted EBITDA was $3.5 billion against $18.4 billion of capex, a net quarterly shortfall of roughly $14.9 billion. The company currently holds about $100 billion of cash, from roughly $75 billion raised in the IPO plus $25 billion of bonds issued since (Q2 earnings call). At the second quarter's burn rate, $100 billion funds a little more than six quarters — while management's compute roadmap runs from today's 1.4 gigawatts to "closer to 10 gigawatts than 5" by the end of 2027 (Musk, Q2 earnings call). Put those two statements side by side and the answer is unambiguous: this spending cycle cannot be completed with the cash on hand. Morgan Stanley, while maintaining a bullish stance, has said outright that free cash flow may not turn positive until 2035 (Benzinga 2026-08).

The starting point on the liability side is in the prospectus: $28.7 billion of debt and finance leases at the end of the first quarter, $60.5 billion of total liabilities, against just $15.9 billion of cash. The post-IPO $100 billion cash balance and the new $25 billion of bonds have given that balance sheet an entirely new skeleton — and permanently installed an interest expense line inside it.

The dilution is already a matter of record, and it belongs in percentage terms on existing shareholders:

The $14.1 billion of "non-cancellable" cloud contracts that has been circulating needs to be checked against what was actually disclosed. The company described that $14.1 billion on the call as non-cancellable agreements, while Fortune's July reporting stated that agreements of this type carry 90-day cancellation clauses exercisable by either side. If customers accelerate their own in-house compute buildouts, the pace at which that $14.1 billion converts is not as rigid as the contract language suggests. This is the single most questionable number in the whole piece.

There is one more piece of buried arithmetic on earnings quality. Second-quarter adjusted EBITDA was $3.5 billion against a GAAP net loss of $541 million, a gap of roughly $4 billion made up of depreciation and amortization, stock-based compensation and interest. That $4 billion is not outrageous today, because most of the $15.8 billion of compute equipment was only installed this quarter and has been depreciated for a few months at most. Once that fleet depreciates at full load — on top of what will be added in the third and fourth quarters — this line will be pushing up by tens of billions per quarter through next year. The CFO's offsetting claim is that "payback on newly deployed compute capital is under a year." If that holds, depreciation is not frightening at all; if compute rental rates drift lower, it is the most expensive sentence in this entire report. Chapter 6 gives a firm view on which way rents are going.

Management has already named the real upstream bottleneck: memory. Global DRAM contract prices rose 80% to 90% sequentially in the first quarter of 2026, with some suppliers telling customers to budget another 10% to 20% per month for the year. IDC expects 2026 DRAM supply to grow only 16%, data centers already absorb roughly 70% of global memory output, and SK Hynix has warned the shortage could persist beyond 2030 (CNBC 2026-01-10 / IDC). The GPUs themselves come from Nvidia, and the Starmind orbital AI program runs on Nvidia hardware too. To pull that link in-house, the company has partnered with Tesla and Intel on the Terafab chip manufacturing complex in East Texas — $16.8 billion for phase one, with Grimes County filings showing a full-program ceiling of up to $119 billion, and the company expecting to absorb roughly 75% of Terafab's output (Simply Wall St / ts2.tech 2026-08). The distinction matters: $16.8 billion is a committed phase one; $119 billion is a planning ceiling, and several years and several financing rounds separate the two. Power is the other hard constraint — going from 1.4 gigawatts to 10 takes more than money.

Customer concentration downstream is the most fragile part of this business today. Anthropic and Google together represent roughly $26 billion of annualized revenue against total AI segment revenue of $2.56 billion in the second quarter — in other words, the growth curve of that segment rests almost entirely on two customers, and both agreements carry 90-day cancellation clauses. Connectivity is far more diversified: 12 million consumer subscribers plus $1.8 billion of quarterly enterprise and government revenue growing 108% year over year is a healthy customer structure. The space segment sells to NASA and the Department of Defense, with Starship carrying a critical piece of the Artemis lunar program — steady revenue, but only 29% growth, and still a $205 million loss.

There are three soft spots in the contract structure: only $14.1 billion of the $47.5 billion backlog is cloud services, with the rest not itemized; that $14.1 billion itself carries cancellation clauses; and Terafab's $119 billion is a ceiling figure. Taking supply and demand together, the weakest link is jointly determined by memory supply and by two AI customers' willingness to renew — the former sets this business's costs, the latter sets its revenue, and neither is currently in the company's own hands.

The cleanest comparison for the AI segment is CoreWeave, because both companies are in exactly the same business: put GPUs in a data hall and rent them by the hour. Line up the two sets of public numbers for the same period and the result is uncomfortable:

For the same revenue, SPCX spent close to twice as much, has one-seventh the contracted backlog, and earns 16 percentage points less per unit. That comparison is not a verdict — a meaningful share of SPCX's compute is for internal use, since Grok and Starmind both consume it, and self-consumed capacity never shows up in backlog by definition. But it does establish that if you value this purely as a business that rents compute to third parties, the visibility SPCX currently offers is clearly inferior to that of a company two orders of magnitude smaller by market value.

The comparison for the space segment is Rocket Lab. In its August 10 disclosure, RKLB reported backlog of $2.36 billion, $1 billion booked in the quarter, backlog up 137% year over year, and issued its first quarterly revenue guidance of $250 million to $265 million. SPCX's space segment did $962 million of revenue in the quarter, against a company-wide backlog of $47.5 billion. The direction of travel is identical and the scale differs by an order of magnitude — but SPCX's segment is still loss-making and RKLB is growing faster. Neither company has reached the harvest phase of the launch business.

One more scale reference: among traditional defense names reporting the same period, HII disclosed a $57.3 billion backlog with $6.7 billion booked in the quarter, and CAE a $10.7 billion backlog (both Q2 2026 disclosures). Set against those coordinates, SPCX's $47.5 billion backlog is not even the largest. The difference is entirely growth rate and composition — a traditional defense backlog is a decade of government orders, while SPCX's $14.1 billion of cloud contracts is a set of commercial agreements that can disappear in 90 days.

The price-to-earnings ruler is useless here. The company is still loss-making on a GAAP basis, and on a non-GAAP basis the 2026 forward P/E is 378.6x (stockanalysis) — a number that tells you nothing except that the denominator is close to zero. The rulers that work are enterprise value to revenue and enterprise value to annualized run rate.

What does 41x mean in practice? It means the market has already paid, as if it were a settled fact, for a second half in which revenue doubles to $32.1 billion — something that has not happened yet. The 18.3x figure looks far more reasonable, on the condition that the $100 billion annualized run rate actually shows up in December. First-half revenue was $12.50 billion; between here and $8.3 billion in the single month of December sit $6.7 billion of newly signed contracts that only begin ramping in October, a Terafab phase one that has not broken ground, and two large customers who can walk away with 90 days' notice.

The spread among sell-side analysts is the best footnote available: the average target across 35 firms is roughly 63% above the current price, the most bullish sees 5.6x the current price, and the most bearish sees 0.44x (S&P Global compilation / stockanalysis). When the same company and the same report produce a range that wide, it means the two camps are not valuing the same business at all — the bulls are discounting a $1 trillion revenue target for 2030, and the bears are discounting a balance sheet that generates no free cash flow before 2035.

This industry is governed by two price curves at once, and right now they are moving in opposite directions.

On price direction, upstream memory is still rising — DRAM contract prices up 80% to 90% sequentially in the first quarter, with suppliers telling customers to budget another 10% to 20% per month (CNBC 2026-01-10 / industry checks). Downstream compute rents are already falling: cloud rental pricing for Nvidia's B200 has dropped from a late-May peak of $6.11 per hour to $4.22, roughly 30% in three weeks, with multiple firms expecting stabilization at $2.50 to $3.00 per hour in the fourth quarter, while B300 spot has already been seen at $2.45 (industry price index, 2026-08). On utilization and scarcity, IDC puts 2026 DRAM supply growth at only 16% while data centers absorb roughly 70% of global memory output — the shortage is real. On the compute side, the three direct causes of the B200 price decline are improving TSMC 4NP yields, materially easier HBM3e supply from SK Hynix and Micron in the second quarter, and more new cloud providers taking delivery and opening spot instances — supply is catching up. On payback, the company's own figure is "under a year for newly deployed compute capital."

The read is this: upstream memory is in the later-middle innings of its upswing, still rising but with the first signs of supply relief; compute rental pricing has already slid out of the end of its upswing into the first few months of a downswing; and the buildout spend itself is still in its earliest phase, with capex still doubling quarter over quarter. The people selling shovels are still making money, the people renting shovels are cutting prices, and the people buying shovels are adding — that is the textbook middle of a capex cycle, and historically the exact position where "payback under a year" stops being true: every 30% decline in rents extends the payback period by a third.

Two leading signals worth tracking. First, cloud spot quotes for B200 and B300 stabilizing and turning up — that would be the first real evidence that demand has caught supply. Second, the three big memory makers publishing firm capacity-expansion and production timelines, or DRAM contract price increases narrowing to single digits sequentially — that would mark the top of this upstream shortage, with cost pressure easing before revenue pressure does.

This chapter is about week-scale signals and governs entry and exit timing over the next few weeks. Readers with a horizon beyond a year can skip straight to the next chapter.

The distribution of money on August 13, 2026 has a clear direction. The three strongest inflows are gold (GLD, strength reading 57), intermediate Treasuries (IEF, 43) and silver (SLV, 38); the three heaviest outflows are retail REITs (SPG, −70), short-term volatility (VIXY, −67) and Mexico (EWW, −63). The overall market state reads bullish with a strength of 43 (on a −100 to +100 scale), while the rotation signal across major asset groups is neutral. The three strongest migrations are utilities into consumer discretionary (strength 4.77), utilities into materials (3.98) and consumer staples into materials (1.85) — money is leaving defensive utilities and staples, but it is landing in materials and discretionary, not in growth. Materials is already confirmed, with gold miners upgraded on August 7 and silver and copper miners on August 10. Healthcare is confirmed as well. Industrials — including defense tech upgraded on August 7 and drones on August 10 — is still building and not yet confirmed.

Translated into a trader's terms: the buy orders this week are concentrated in gold and Treasuries, and this stock's sector is not on any confirmed money-flow leadership line (2026-08-13). The macro backdrop is no help either — net liquidity is $5.84 trillion, up just 0.3% week over week and essentially flat, with the 10-year Treasury yield at 4.7%, leaving liquidity on the tighter side.

Has price run ahead of the money? The stock rose 21.1% in the week of August 10, including a 15.83% single-day gain on August 7 on 236.7 million shares, 91% above the three-month average volume (CNBC / industry reports 2026-08). Two things drove that move: the large second-quarter revenue beat, and the absence of panic selling after roughly 911.5 million shares unlocked on August 6. That is short covering and sentiment repair as an overhang cleared, not new allocation money building positions — and the flow picture above shows incremental money did not come this direction at all this week.

On the long-horizon line, the answer is that this is worth following, but only after you have priced the cost of following it. On fundamentals, 92% revenue growth, 213% sequential growth in the AI segment with its first profitable quarter, and 108% growth in enterprise and government connectivity revenue are all genuine operating improvements. On valuation, 41x 2026 revenue for a company growing close to 100% is not absurd, and 18.3x the annualized run rate is arguably cheap. There are three costs to accept. Free cash flow, on even the most optimistic sell-side framing, waits until 2035, and several more equity or debt raises are almost certain in between. Of the dilution already committed, the 3.4% from Cursor plus the stock portion of the spectrum consideration is on its way. And the share supply overhang is not truly cleared until Musk's roughly 6.4 billion shares unlock on June 12, 2027. Position sizing should be driven by verification points, not by the share price — each quarter, you only need to confirm three things: that AI segment revenue is still growing at a multiple, that quarterly capex growth has begun to run slower than revenue growth, and that the cloud contract figure is expanding rather than being cancelled. All three hold, keep holding; one breaks, redo the math.

On the short-horizon line, the answer is do not chase. Money leadership right now sits in gold, Treasuries and materials, and this stock's sector has not even reached "confirmed." The share price has just returned to roughly its average since listing, with the entire trapped-supply band from June's sentiment peak overhead and wave after wave of unlock supply below — several day-count tranches land between September and October, and the early-November earnings report itself triggers a larger one. What to wait for? Either money rotating out of the defensive side, or price holding above the mean line with volume falling back to normal — that is what it looks like when incremental money genuinely starts taking the other side.

The specific conditions and the specific levels follow below.

Two entry conditions: one watches the money, one watches the numbers.

Condition one (money flow, trackable daily): industrials moves from "money is gathering" to a confirmed leadership line, while the strength readings for gold and intermediate Treasuries fall meaningfully from the August 13 levels of 57 and 43. Those two happening together means defensive allocation is loosening its grip and growth and manufacturing are receiving the money — only then does this stock have incremental buying to speak of. Baseline date is 2026-08-13.

Condition two (the next quarterly report, derived from the math thresholds above): third-quarter results (expected after the close on November 3, 2026, per Wall Street Horizon's estimate; the date is not confirmed by the company) need revenue at or above $12.0 billion, AI segment revenue above $4.0 billion for the quarter, and quarterly capital expenditure no higher than $20.0 billion. All three together are what keeps the $44.63 billion full-year consensus intact; if revenue comes in below $12.0 billion while capex keeps expanding, the valuation denominator and the cash burn both move the wrong way at once.

Three reference levels, each anchored to something publicly verifiable:

The invalidation level is about 26% below the current price, which leaves a lot of empty space in between, so pair it with an early warning: a close below $135 that is not recovered above the 50-day moving average ($141.86 as of 2026-08-13) within two weeks should be treated as thesis damage, without waiting for $104.83.

For the record on where the stock sits now: the current price is $142.27, down 2.7% on the day, sitting just above both the 50-day and 200-day moving averages — with only two months of trading history, those two lines are currently the same number ($141.86), so that level simultaneously represents the short-term mean and the entire average cost basis of everyone who has traded the stock publicly. On the sell side, the average target across 35 firms is $231.4, with a high of $800 and a low of $62, and Morgan Stanley maintains overweight with a $300 target (S&P Global compilation / stockanalysis / 2026-08).

Two things to watch daily: where the close sits relative to $141.86 — three consecutive sessions holding above it without a high-volume drop says willingness to hold above the mean is recovering; and whether the strength readings for gold and Treasuries are coming down from 57 and 43, with a fall below 30 the point at which defensive sentiment has genuinely loosened.

Two things to watch weekly: cloud spot rental quotes for B200 and B300 — stabilizing and turning up is the bottoming signal from chapter 6, while continuing to slide toward $2.50 per hour means the compute payback period is stretching; and the sequential increase in DRAM contract prices, where narrowing to single digits means cost-side pressure has peaked.

Five things to watch each quarter: whether sequential revenue reaches the $12.0 billion line; AI segment revenue and adjusted EBITDA margin (against CoreWeave's 59%); the absolute level and sequential direction of quarterly capex; whether cloud services contracts move up from $14.1 billion or get cancelled; and the gap between adjusted EBITDA and the GAAP net loss — how fast that gap widens from the roughly $4 billion of the second quarter is the speed at which depreciation is coming through.

We will publish a supplement if any of the following occurs: actual dilution from the Cursor close exceeds 3.4%; either cloud customer exercises its 90-day cancellation right; quarterly capex breaks $20.0 billion while revenue fails to reach $12.0 billion; an unlock date produces a high-volume break below $135; or the company announces a new equity raise.

Catalyst calendar:

Every number in that tracking list is something we watch daily — whether money rotates out of gold and Treasuries, whether compute rents stop falling, whether DRAM contract price increases narrow. If any of them triggers, we publish a supplement alongside that day's daily review, so you do not have to sit on the data yourself. Every money-flow, sector-strength and macro-liquidity reading that appears in this article can be checked line by line on the corresponding data pages of the site, with dates matching those cited here; for the company-level revenue, spending, contract and unlock schedule figures, the sources are in parentheses and can be verified against the original disclosures. The third-quarter report will turn the $12.0 billion threshold from chapter 1 into a definite answer. Until then, the conditions and the levels are laid out above.

This article was produced on the NextPick research workbench; every money-flow, tier and valuation figure it cites can be verified on the corresponding data surfaces of the site, and the moment any tracking condition listed in the report triggers, an addendum will go out together with that day's daily review.