Newsletter · · Ashutosh Agarwal

SpaceX Has No Profit Engine to Fund Its AI Buildout - The AI Capex Tracker - Week of August 6, 2026

A synthesis of what investor podcasts said about AI capital spending around August 5 to 6, 2026, as SpaceX's first earnings report showed $18.4 billion of spending against $7.8 billion of revenue and became the cleanest test yet of whether a giant AI build can be funded without a profitable business underneath, for the week of August 6, 2026.

The AI Capex Tracker

Week of August 6, 2026: SpaceX Has No Profit Engine to Fund Its AI Buildout

TL;DR

  • SpaceX's first-ever earnings report is the cleanest test yet of the question hanging over the whole trade: can you fund a giant AI build without a profitable business underneath it? Second-quarter revenue was $7.8 billion, but the company spent $18.4 billion building things, most of it on AI data centers. That is more than twice what it takes in. The only profitable division is Starlink. The stock fell 7% after the report and is now down more than 40% since its June IPO, erasing over $1 trillion in value. (Morning Brew Daily, Aug 5; Bloomberg Intelligence, Aug 5)

  • Goldman Sachs's own credit strategists just put hard numbers on how much of this build is going onto the bond market, and where the buyers start saying no. They expect roughly a third of hyperscaler capex to be paid for with borrowed money in 2026 (about $250 billion of new bonds, on top of $194 billion already issued this year), peaking at 35% in 2027. The tell that demand is tiring: the number of big insurance buyers placing large orders for long-dated AI bonds roughly halved between the first and second quarter. (Exchanges, Aug 5)

  • A Sequoia partner who has tracked this math for two years says the industry now needs about $4 trillion of lifetime revenue to pay back what it is spending, and only true "AGI" would justify it. David Cahn's blunt read on 2026: "I think we reached a point where people just don't care about the math anymore at all. It's all in, all the way." Meanwhile the lab bosses themselves keep saying human-level AI is 10 to 20 years away. (Big Technology Podcast, Aug 5)

What's new

A note on the window: this is a Thursday issue, so it sweeps roughly the last 24 hours, podcasts published August 5 and 6, with a couple of late August 4 episodes that had not shown up before. Yesterday's issue was about who is buying the AI (two loss-making labs, per named sell-side estimates). Today the conversation moves one link down the chain: who can pay for the build. A newly public SpaceX became the live case study, Goldman's credit desk quantified how much of the build is being borrowed, a serious AI bull spelled out the size of the bill, and, out of nowhere, Texas hit the brakes. Ranked by where the dollars of profit-and-loss impact are biggest.

1. SpaceX reported for the first time as a public company, and it is the purest example of the "can you afford this?" problem. Morning Brew Daily, Aug 5 (hosts Neil Freiman and Toby Howell) and Bloomberg Intelligence, Aug 5 (analysts Mandeep Singh and Ed Ludlow). The first is a general news show; the second is Bloomberg's own equity analysts, so treat their modeling as the more rigorous read.

The numbers, plainly: quarterly revenue came in at $7.8 billion, beating expectations by more than a billion. But capital spending, the money laid out to build things, was $18.4 billion, "largely driven by its AI data center expansion." As the hosts put it, the company is "spending more than twice what it makes in quarterly revenue." Elon Musk repeated a goal of $1 trillion in annual revenue by 2030 (a timeline he says moved up a year), and the finance chief said the run rate should hit a $100 billion annual pace by year-end. But of the three segments, only one makes money: Connectivity (Starlink) at $4.3 billion; Space lost money at $962 million, and the AI arm lost money at $2.56 billion.

Bloomberg's Mandeep Singh did the cash math with co-host Ed Ludlow live on air:

"In the first half of the year, SpaceX had about $3.5 billion of operating cash flow. But the capital expenditures for the first half of the year were like $29 billion… let's say $25 billion of negative free cash flow in aggregate."

Free cash flow is simply the cash a business has left after paying to run and build itself; SpaceX's is deeply negative, potentially near -$50 billion for the full year on Singh's annualized math. And that is the whole point. As Morning Brew framed it, Microsoft, Meta and Amazon got punished for their AI spending too, "but the thing that they have that SpaceX doesn't is that they're funding it using very profitable existing businesses." Singh's conclusion: "there is a ceiling to how much higher they can go in terms of CapEx increases… how high can they go given they don't have the balance sheet and the free cash flow that the other hyperscalers have?" The stock, up 9% into the print, fell about 7% after it, and is down over 40% from its IPO, more than $1 trillion of market value gone. A separate pressure point landed the same day: the post-IPO lockup expired, freeing insiders to sell up to 20% of shares (about $100 billion worth), which the hosts flagged as a near-term overhang.

Why this moves the thesis: SpaceX strips the argument down to its bones. Every hyperscaler bull case quietly rests on "the core business pays for it." SpaceX doesn't have that cushion, the market repriced it hard, and it is now the cautionary tale every skeptic will point to.

2. There is a real, new Nvidia read-through buried in the SpaceX story: an exclusive chip commitment and a plan for data centers in orbit. Morning Brew Daily, Aug 5; CNBC's Fast Money, Aug 4 (reporter Morgan Brennan on the call); Squawk Box Europe Express, Aug 5 (former Tesla board member Steve Wesley).

On the call, SpaceX said it will build all of its AI infrastructure exclusively on Nvidia chips, and announced a joint venture with Nvidia to design a satellite compute payload ("Starmind AI-1") to enable orbital data centers, with first launches teased as soon as next year. Morgan Brennan, reporting live, added the scale: SpaceX expects to end this year with over 2 gigawatts of compute, and Musk framed next year as "closer to 10 gigawatts than five," with the company "looking at 15 gigawatts of power plant at the power plant level for next year." (A gigawatt is roughly the output of one large nuclear reactor, so we are talking about a single company chasing the power of a dozen-plus reactors.) Steve Wesley's read on why Musk keeps putting Nvidia front and center: in a world short of chips, "if you're a company who has access to chips… that is a very strong competitive advantage."

Why it matters: this is another marquee, all-in Nvidia customer locking itself to Nvidia silicon just weeks before Nvidia's own quarter, a small but real demand signal, and a reminder that even the newest entrant isn't designing its way around Nvidia.

3. Goldman's credit desk quantified how much of the build is being borrowed, and where the buyers push back. Exchanges, Aug 5, Goldman Sachs credit strategist Amanda Lynam (who leads Credit Strategy Research) and Zach Ablon (credit sales desk), hosted by Alison Nathan. This is exactly the kind of primary, desk-level source that beats punditry, and it is the most important financing read of the cycle.

The headline framing: because capex estimates keep getting revised up, Goldman measures AI borrowing as a share of capex rather than a fixed dollar figure. Last year (2025), hyperscalers issued $108 billion of bonds, about 27% of their capex. This year they have already issued $194 billion, and Lynam expects roughly one-third (33%) of capex to be debt-financed, or about $250 billion of new hyperscaler bonds. She expects that to peak in 2027 at 35%, precisely because "CapEx and cash flow from operations are quickly approaching the same level, they're converging", in plain terms, the companies are running out of spare internal cash and must lean on the bond market. To size the shift: AI-related issuance was 1% of all investment-grade bond supply in 2024, ~7% in 2025, and about 18% this year, and 40% of all long-dated (15-year-plus) investment-grade issuance this year has been AI-related. Amazon is now the single largest long-duration name in the index (it was 20th last year); Google jumped to 18th from 86th.

The nuance, and the sell signal, is where demand tires. Goldman is emphatic that access to capital is not the problem (hyperscalers could add ~$2 trillion of debt and stay investment-grade; the US market could absorb maybe $510 billion more before hitting bank-like concentration limits; private markets hold ~$4.5 trillion of dry powder). The problem is price and appetite. Ablon's desk color is the tell:

"When we take a look at the last 12 months for our AI leader basket, it's gone from tights of 74 basis points to nearly twice that. So the marketplace has clearly expressed some indigestion."

Even sharper: the big insurance buyers ("real money") who anchor 30-year issuance were placing roughly 15 large ($50 million-plus) orders on AI deals early in the year; by late in the second quarter "that number was about half as big… clearly losing their appetite for this risk." So the bond market isn't shut, but the marginal buyer of the longest, riskiest AI paper is stepping back, and spreads are widening. That is the mechanism the bears have been describing, now measured.

4. Out of nowhere: Texas halted new data-center hookups to its grid. Energy News Beat Podcast, Aug 5, host Stuart Turley (energy commentator).

On August 3, Governor Greg Abbott directed a halt on new data-center connections to the ERCOT grid (the Texas power system) pending a full audit of their energy use, water consumption, financing, ownership, and community impact, instructing the state utility commission and ERCOT to verify tax incentives, grants and projected income. Abbott's stated priority: "providing affordable energy, maintaining a reliable electric grid and preserving water for our communities." Turley put the scale of what is now under review in context: ERCOT is tracking "more than 1,800 projects in its interconnection queue… 474 gigawatts of requested capacity." The move came about three weeks after New York became the first state to impose a temporary moratorium on new hyperscaler data centers (though New York's involved only five projects).

Why it matters: Texas has been the answer to the power bottleneck, the prior issue cited estimates of Texas grid demand climbing from 85 to 150 gigawatts by 2030. A governor slowing new connections in the single most important build-out state is a genuine new risk to the timeline, and a reminder that the binding constraint on this trade increasingly is politics and physics, not chips. Note the source here is an opinionated solo commentator; the underlying facts (the Abbott directive, the queue size, the New York moratorium) are the citable part, not his editorializing.

5. The "AI capex illusion": Google has quietly committed ~$700 billion to power, and depreciation hasn't even started biting. The Investing for Beginners Podcast, Aug 6, guest Thomas Chua of Steady Compounding (this was the single freshest episode of the sweep, published today).

Chua's point is about the numbers that don't make the headlines. Reading Alphabet's second-quarter filing, he found that Google's long-term purchase commitments spiked from a couple of billion dollars in a typical quarter to "over 700 billion dollars" this quarter (in note 10 of the filing), largely multi-decade contracts for electricity to power data centers, which the company "can't really get out of… unless they pay a heavy penalty." His broader warning is an accounting one. Depreciation is how a company spreads the cost of an expensive asset over its useful life, lowering reported profit each year. But half-built data-center gear sits in a balance-sheet line called "construction in progress", and while it sits there, it is not yet being depreciated:

"It means they have the raw ingredients to create all these data centers, but… it's not subjected to depreciation, which is why we still see margins increasing tremendously for a lot of these hyperscalers."

In other words, some of today's fat margins are flattered because the depreciation bill on all this new gear hasn't landed yet. Chua's read on winners and losers: the big diversified players (Amazon, Microsoft, Alphabet) can absorb excess compute in their own core businesses if outside demand softens; the newer, debt-heavy cloud providers and Oracle, "leveled up to their eyeballs," with core businesses that are "just purely selling compute", are far more exposed. He also relayed a striking demand data point: Meta's finance chief Susan Lee said the company is fielding offers to buy its compute at "multiples of what they paid for."

The debate

Steel-manning both sides of the ~$1.5-trillion-in-2026 (and larger in 2027) AI capex thesis. Last issue the fight was demand vs. debt. This cycle it sharpened into something even more fundamental: is the spending justified because the payoff (and the financing) is real and coming, or is the industry spending on faith, against a bill it may never clear?

Bull, the demand is real, the returns are showing, and the money is available. Google's cloud revenue grew 82% last quarter to $24.8 billion with a $514 billion backlog of signed multi-year contracts, and corporate token usage is reportedly up around 1,000% this year (Practical News, Aug 5; Trader Talk, Aug 5). Meta's ad business is reaccelerating above 20% while it fields offers to rent its compute at "multiples" of cost (The Investing for Beginners Podcast, Aug 6). And on financing, Goldman's own credit team says flatly it is "not concerned about access to capital", the hyperscalers are under-levered, and between the bond market, private credit's ~$4.5 trillion of dry powder, and converging cash flow, the build can be funded (Exchanges, Aug 5). The bull's clincher is still last week's Gavin Baker point: the price of renting a GPU is rising, so the owners are under-earning and cash flow should accelerate.

Bear, the bill is measured in trillions, and only a technology that doesn't exist yet would justify it. David Cahn, a Sequoia partner who has been an AI bull for a decade, laid out the arithmetic that gives even bulls pause. His rule of thumb: for every dollar of GPU capex, roughly a dollar of energy is needed to run it, and the builder needs a ~50% margin, so each year of GPU spending requires about twice its value in lifetime revenue to pay back. Stacking the years since ChatGPT, his "$X billion question" series ran $200 billion (2023), $600 billion (2024), $850 billion (2025), and about $1.5 trillion (2026), he gets "about $3 trillion that needs to get paid back just since ChatGPT," rising past $4 trillion once 2027 is added. His verdict: "the only possible way to pay those dollars back is going to be AGI", human-level artificial intelligence, and yet the people building it (Altman, Ilya Sutskever, Dario Amodei) keep saying AGI is "10 or 20 years away." The mismatch, and the danger, is that markets are pricing near-term payback the labs themselves aren't promising (Big Technology Podcast, Aug 5). The macro voices layer on the timing risk: Man Group's Christina Hubu warns the setup "rhymes" with the late-1990s telecom bubble and that hyperscalers face a Bain forecast requiring "a very high level of AI-related revenues" by 2030 that will be "a stretch," while Legato's Ryan Kelly notes Google is now spending more capex than it generates in cash "first time ever" and that GPUs "can easily become obsolete… because more powerful GPUs are created" (Trader Talk, Aug 5). And a CNBC trader panel flagged "$1.65 trillion of private debt that is associated with this trade", much of it off balance sheet, like Blackstone's ~$30 billion facility lending Anthropic money to buy Google TPUs with the chips themselves as collateral, warning it "unravels the same way in 2000" if rates rise or demand pulls back (CNBC's Fast Money, Aug 4).

The synthesis. The two sides are no longer really arguing about demand, even the bears concede token usage and cloud backlogs are exploding. They are arguing about time and money: whether the revenue and the financing arrive before the depreciation, the debt-service, and investor patience run out. Cahn's framing is the honest one, a bifurcated path where either AGI shows up and pays for everything, or "we got the timing wrong… and you have a big reckoning that has to come." SpaceX is the live rehearsal of that reckoning for any name without a profit engine. And Goldman's desk gives you the dashboard: the money is there, but the price is rising and the marginal buyer is stepping back. Watch the price of credit, not just the availability of it.

Sell signals to watch: - AI credit spreads widening further. Goldman's AI basket already went from 74 basis points to nearly double that; if it keeps widening, and if the big insurance buyers keep pulling back from long-dated AI bonds, the financing cost that underwrites the whole build is rising in real time. - The Texas halt spreading or sticking. If Abbott's audit stalls a meaningful slice of the 474-gigawatt ERCOT queue, or if more states follow New York and Texas, the power bottleneck stops being a 2027-2028 worry and becomes a right-now delay to committed builds. - A weak SpaceX aftermath post-lockup. With ~$100 billion of stock freed to sell, a heavy insider exit would signal that even believers want out at these levels, and would make the "no profit engine" names harder to finance. - Capex outrunning cash flow at more hyperscalers. Google already crossed into negative free cash flow. If a second or third hyperscaler follows, the "self-funding" bull case weakens and the debt share climbs past Goldman's 35% peak estimate. - Depreciation finally landing. As "construction in progress" converts to live, depreciating assets, reported margins should compress. The first quarter a hyperscaler's margin visibly bends under depreciation is a real thesis test.

Stocks in play

NVDA. Bull: fresh, concrete demand, SpaceX committed to build all its AI infrastructure exclusively on Nvidia chips and formed an orbital-compute joint venture with Nvidia, one more marquee, all-in customer weeks before Nvidia's own quarter (Morning Brew Daily, Aug 5; CNBC's Fast Money, Aug 4). Bear: it sits at the center of every off-balance-sheet financing scheme (chips and TPUs pledged as loan collateral) and every "who actually pays?" question, if the credit or demand air pockets, Nvidia's order book is the first thing to wobble (CNBC's Fast Money, Aug 4). Next: its late-August quarter, still the single demand read for the entire complex.

AVGO. Bull: the "own the chip" custom-silicon theme is intact, hyperscalers keep diversifying away from Nvidia, and every custom-ASIC and TPU deal (including chips being used as loan collateral) validates Broadcom's role. Bear: no standalone Broadcom operator signal this cycle, so it trades on the complex's beta into a jumpier credit backdrop. Next: read-through from the hyperscalers' custom-chip ramps; otherwise quiet in the podcasts this cycle.

AMD. Bull: the "credible second source" story got louder, AMD has now reportedly landed all five major frontier AI labs on a single rack platform, with deals "measured in gigawatts," plus a new Venice X data-center CPU to pair with Helios in 2027 and Lisa Su's claim the AI-accelerator market reaches $1.4 trillion by 2030 (Practical News, Aug 5). Bear: still up ~140% year-to-date into the print, with third-quarter gross margin guided to 56% (a touch below the ~56.2% hoped for), and the big Helios/MI450 revenue is a late-2026-into-2027 story, not a today story (Bloomberg Intelligence, Aug 5). Next: Helios/MI450 shipment timing and how much revenue the multi-gigawatt lab deals add in 2027.

MSFT. Bull: the market rewarded it after earnings because management (Satya Nadella) communicated the cloud economics clearly, and it remains diversified enough to absorb its own compute if outside demand softens; it also has ample room to fund the build with cheap debt given how under-levered it is (The Investing for Beginners Podcast, Aug 6; Exchanges, Aug 5). Bear: it still carries the concentration flag from last issue (a large share of its AI revenue tied to two loss-making labs), and having been "beaten up" to ~$350 it now needs Copilot revenue to visibly scale into the spend (Trader Talk, Aug 5). Next: whether AI revenue growth keeps pace with capex, and the promised 2027 free-cash-flow inflection.

GOOGL. Bull: the clearest "returns are showing" story, cloud revenue up 82% to $24.8 billion, a $514 billion contract backlog, Gemini at 950 million monthly users, and quarterly profit of $112.1 billion (about 4x a year ago) (Practical News, Aug 5). Bear: it raised capex to $205 billion and slipped into negative free cash flow "first time ever," and it has quietly committed roughly $700 billion to long-term power contracts it "can't really get out of," with margins still flattered by depreciation that hasn't started (Trader Talk, Aug 5; The Investing for Beginners Podcast, Aug 6). Next: whether core profit keeps growing into the raised 2027 spend, and how those footnote power commitments convert to real cost.

AMZN. Bull: the market rewarded it post-earnings for clear communication from Andy Jassy on cloud economics, and it has the most diversified core to absorb any compute glut (The Investing for Beginners Podcast, Aug 6). Bear: it is now the single highest long-duration bond weight in the investment-grade index (up from 20th a year ago), a sign of just how much it is borrowing to build, and it still carries last issue's concentration flag (an estimated 73% of its AI revenue tied to OpenAI and Anthropic) (Exchanges, Aug 5). Next: whether AWS growth keeps outpacing the capex raise as the debt load climbs.

META. Bull: the core is humming, advertising reaccelerating above 20% with double-digit engagement growth, and it is fielding offers to rent its compute at "multiples of what they paid for" (The Investing for Beginners Podcast, Aug 6). Bear: it was punished after earnings because management could not clearly explain the return on the huge compute spend at its Superintelligence Lab, and one trader panel argued it may "no longer" deserve a premium multiple now that free cash flow has collapsed under the build (The Investing for Beginners Podcast, Aug 6; CNBC's Fast Money, Aug 4). Next: an actual monetization plan for the Superintelligence Lab compute, still the missing piece.

Read-throughs

  • Power / grid, the constraint just became a policy event. The single most actionable power development of the cycle is Governor Abbott's August 3 halt on new data-center connections to the Texas grid, pending an audit of energy and water use, against an ERCOT queue of 1,800+ projects and 474 gigawatts of requested capacity (Energy News Beat, Aug 5). The read-through is two-sided: a near-term headwind for anyone whose build timeline runs through Texas, but a longer-term tailwind for behind-the-meter power (gas turbines, on-site generation) that lets a data center sidestep the public grid entirely. Morgan Stanley's own podcast framed the macro version of the same wall, a 38-gigawatt US power shortfall through 2028 and a push toward behind-the-meter generation and better capacity utilization (today just 30-40%) as the practical fixes (Thoughts on the Market, Aug 4).

  • The grid price signal keeps rising. In the PJM market (the mid-Atlantic grid), the latest capacity auction cleared at the $325/MW-day federal price cap, but would have cleared at $554.72 without the cap, and 23 gigawatts of capacity chose not to bid at all, per LS Power's CEO (Energy Evolution, Aug 4). Translation: the true clearing price of power is being suppressed by caps, real scarcity is worse than the headline, and that is a structural tailwind for merchant power generators and anyone selling firm capacity.

  • Memory, still the cheap way to play the build, if the demand holds. No fresh operator pricing print this cycle, but a retail-investor show flagged the setup that keeps drawing attention: Micron trading around a 6x forward earnings multiple, a ~70% discount to the S&P 500, against memory demand said to be growing far faster than supply (Buy Hold Rant, Aug 5). The structural bid from last issue (more memory per chip means more tokens out) is intact; the risk is the same as the whole trade, a demand air pocket.

What changed vs last issue

Last issue (Wednesday, Aug 5, "73% of Amazon's AI revenue: two loss-making customers.") was about who is buying the AI, named sell-side estimates that a huge share of hyperscaler AI revenue traces to OpenAI and Anthropic, plus Gavin Baker's bull rebuttal that rising GPU rental prices make the build self-funding. Today's ~24-hour sweep moved the story to who can pay for the build and put hard numbers on the financing.

  • A brand-new, vivid test case: SpaceX. Last issue had no SpaceX. This issue it reported for the first time, $7.8 billion revenue against $18.4 billion of capex, roughly $25 billion of negative free cash flow in the first half, only Starlink profitable, and the stock down 40%+ from its IPO. It is the purest illustration of the "no profit engine" risk, and it dragged an exclusive Nvidia chip commitment and an orbital-data-center JV into view.

  • The financing worry got quantified by Goldman's own desk. Last issue's bear case on credit was Jesse Felder's macro warning. This issue has Goldman Sachs credit strategists: 33% of 2026 hyperscaler capex debt-financed ($250 billion of new bonds; $194 billion already issued), peaking at 35% in 2027; AI now ~18% of all investment-grade supply (up from 1% in 2024); the AI credit spread nearly doubling off its tights; and big insurance buyers of long-dated AI bonds roughly halving between Q1 and Q2.

  • The size of the bill got a number: ~$4 trillion. Last issue debated demand concentration. This issue has Sequoia's David Cahn stacking ~$3 trillion of cumulative capex since ChatGPT (rising past $4 trillion of required lifetime revenue with 2027), arguing "nothing short of AGI" justifies it while the labs themselves say AGI is 10-20 years out.

  • Power flipped from a physics story to a policy shock. Last issue's power item was Texas demand climbing 85→150 GW by 2030. This issue: Governor Abbott halted new Texas grid connections on August 3, pending an audit, against an 1,800-project, 474-GW queue, three weeks after New York's moratorium.

  • A new accounting angle: the depreciation illusion. Thomas Chua flagged Google's ~$700 billion of long-term power commitments buried in the footnotes and the "construction in progress" assets not yet being depreciated, a reason today's fat margins may be temporarily flattered.

  • Fresh, concrete data points: SpaceX Q2 revenue $7.8B vs capex $18.4B (AI-segment capex $15.83B vs $13.09B expected), H1 free cash flow ~-$25B, backlog $47.5B, stock -7% post-print and -40%+ since IPO (>$1T erased), lockup freeing ~$100B of stock; Goldman: 33% of 2026 capex debt-financed (~$250B), 35% peak in 2027, $194B issued YTD, AI = ~18% of IG supply and 40% of 15yr+ issuance, AI basket spread 74bp → ~2x, insurance orders halved Q1→Q2; Cahn's ~$3T cumulative / ~$4T lifetime-revenue math; Abbott halt vs 1,800 projects / 474 GW ERCOT queue; Google $700B power commitments, capex $205B, cloud +82% to $24.8B, backlog $514B, Gemini 950M users, profit $112.1B; AMD five frontier labs on one rack, Venice X (2027), $1.4T accelerator TAM by 2030; PJM auction $325 cap vs $554.72 uncapped, 23 GW did not bid; Micron ~6x forward P/E.

  • Still open: whether SpaceX insiders dump stock now the lockup is off; whether AI credit spreads keep widening and real-money buyers keep retreating; whether the Texas halt spreads or sticks; and whether Nvidia's late-August quarter confirms demand for the whole complex.

Next catalysts to watch: Nvidia's late-August quarter (the demand read for everything upstream); the SpaceX post-lockup insider-selling print in the days ahead; the Abbott/ERCOT audit timeline and any copycat state actions; and AI credit spreads plus the next big long-dated hyperscaler bond deal, the financing market is now the swing factor.