Our Q3 2026 earnings report is now public → https://t.co/IRQs4EDBsS
The live webcast with @ElonMusk, @Gwynne_Shotwell, and @BretWJ will begin in 30 minutes
Anthropic is generating breakneck revenue growth but faces risks that include a narrow customer base and uncertain access to computing power as it prepares for an IPO, according to new disclosures https://t.co/MfRsxINTJq
If you hold AI stocks heading into Tuesday and Wednesday night, watch the big spenders' capex guides first: three of the six companies reporting put 98% or more of last quarter's operating cash flow into capex (SpaceX spent 7.6x what its business brought in, Tesla 1.2x, Meta 0.98x).
The three selling into the buildout (Lam, TI and IBM) spent just 13 to 19 cents of every operating dollar, so their reports are a clean read on orders. Below is the one number per company I'd count as a miss.
The ratio is April–June capex over April–June operating cash flow, from each company's own cash-flow statement. SpaceX only listed in June, so its quarter is the first-half 10-Q minus the Q1 numbers in the IPO prospectus: $18.4B of capex against $2.4B of operating cash.
SpaceX (Tue 10/20, 1:30pm PT webcast)
No numeric guidance to grade against, so I'm watching AI capex divided by AI revenue. It came down from 9.4x in Q1 to 6.2x in Q2 ($15.8B of AI capex on $2.6B of AI revenue). Back above 6.2x means the build is outrunning sales again.
Tesla (Wed 10/21, 2:30pm PT Q&A)
On the Q2 call, management said 2026 capex would top $25B. It spent $8.3B in the first half, so getting there takes about $8.4B a quarter if you split the rest evenly. That even split is my assumption, not company guidance. Q3 capex under $8.4B, or a walk-back on the full-year number, means the plan slipped.
Meta (Wed 10/21, 1:30pm PT call)
Guided 2026 capex to $130–145B, narrowed in July from $125–145B, and Q3 revenue to $61–64B. A cut below the $130B floor is the headline that would rattle the whole supply chain. Revenue under $61B makes that pace of spending much harder to defend.
Lam (Wed 10/21, 2:00pm PT call)
Guided September-quarter revenue to $8.1B ± $0.4B, up from $6.72B in June. Anything under $7.7B, the bottom of its own range, is a miss.
TI (Wed 10/21, 1:30pm PT webcast)
Guided Q3 revenue to $5.65–6.15B after $5.46B in Q2. Under $5.65B is a miss.
IBM (Wed 10/21, 2:00pm PT)
Already trimmed 2026 constant-currency growth to 4–5% in July, from more than 5% in January. A second cut, below 4%, is the miss.
My read: the ratio splits the week cleanly. Meta is funding the build with almost nothing left over ($784M of free cash flow), Tesla spent $1.1B more than its operating cash flow, and SpaceX is covering a much wider gap with the cash it raised in June. That makes their capex guides the first place a slowdown would show up. The sellers barely spend, so their prints tell you whether the orders are real, and for Lam and TI the next-quarter guide counts as much as the print. If Meta holds $130B and Lam clears $7.7B, the buildout has another quarter in it. A held floor is good news for the chip suppliers; for Meta holders it only works if revenue clears $61B too.
Not covered: Microsoft, Alphabet and Amazon report next week, and Oracle later.
Which line breaks first this week: Meta's $130B floor, Lam's $7.7B, or neither?
Sources: SpaceX 10-Q and IPO prospectus; Tesla, Meta, Lam, TI and IBM Q2 releases and filings; Tesla Q2 call; company IR event pages. The ratios and the Tesla pace are my own math.
@McnallieM@AnthropicAI@nscale Check the fine print on the first deal. In its September S-1, Nscale says the Aug 25 Anthropic agreements (up to ~$44.6B, Vera Rubin NVL72 at Monarch) require it to use best efforts to line up financing, and as of filing it had no binding commitments for any of it.
@GlobalMktObserv SOXL closed at $164.27 on Oct 5 and trades near $99.5 today, down about 39%. Unlevered SOXX is down about 15% over the same stretch. From here, SOXL needs roughly +65% just to get back to that close, while SOXX needs about +17%.
@SingularityRes Lam's bar tomorrow is its July guide: $8.1B ± $0.4B for the Sept quarter, 20%+ above June's record $6.72B, with 2026 WFE raised to the low $150Bs. Under $7.7B misses its own numbers. A top-end print with a softer December guide would be the first sign tool orders care.
@SmallCapSnipa Look at where the parent's debt trades, though: CoreWeave's 2031 notes yielded about 15.8% on Oct 15 (CNBC). That ~16% residual in JPM's example only reaches the parent after the asset-level debt is paid off, and that's the layer credit investors are pricing like high yield.
Agreed, easily the more interesting release. One catch from the deployment Q&A: the smallest efficient unit takes 64 GPUs for prefill and 64 for decode, so 128 cards just to run it well. KV cache store and cross-replica load balancing aren't in this release either. Great reference code, but most small teams will still run V4.1 Flash on vLLM or SGLang for now.