Newsletter · · Ashutosh Agarwal

AI Capex Forecasts Keep Rising as Spending Drives 45 Percent of GDP Growth - The AI Capex Tracker - Week of July 16, 2026

The AI Capex Tracker for the week of July 16, 2026. Hyperscaler data-center spending forecasts were revised up to 740 billion dollars for 2026 even through Monday's selloff, with AI now driving roughly 45 percent of US GDP growth.

The AI Capex Tracker

Week of July 16, 2026: AI Capex Forecasts Keep Rising as Spending Drives 45 Percent of GDP Growth

TL;DR

  • The market sold the AI trade on Monday, and then the spending forecasts went up again. Carson Group's team walked through the math: at the start of the year they figured the five big spenders (Microsoft, Alphabet, Amazon, Meta, Oracle) would lay out about $470 billion on data centers in 2026. After two earnings seasons, that number is now $740 billion, from 1.7% to 2.5% of the entire U.S. economy, and the 2027 estimate has jumped from $530 billion to nearly $900 billion. The kicker: about 45% of America's economic growth over the last five quarters came from AI-related spending. The whole economy is now leaning on this one trade. (Facts vs Feelings, Jul 15)
  • The one genuinely new bear signal, "people are paying less for AI", got defused by the exact person who invented the measure. Steve Martin of Silicon Data, whose token-spending index the bears have been quoting, went on the record: yes, his index is down about 20% since May, but it's not demand collapsing, it's customers rationally shopping down to cheaper models for simple tasks, the same way you'd pick a generic drug when you're paying out of pocket. Meanwhile the rental price of even three-year-old chips is still climbing. (Full Signal, Jul 15)
  • And there's a lot more money still waiting to be spent, which is the bull case and the risk. On a16z, a data-center analyst laid out the runway: hyperscalers can still grow spending 20-30% next year, CoreWeave and Oracle can raise far more through debt and equity markets, Brookfield and Blackstone are pivoting hard into AI, and the giant sovereign wealth funds (Abu Dhabi's G42, Norway, Singapore's GIC) "have barely started." His warning is the tell: a lot of this money "will [be spent] because they believe," not because a spreadsheet says it pays. (The a16z Show, Jul 15)

What's new

A quick word on the window. This is a Thursday recap, sweeping roughly the last day (Tuesday–Wednesday, July 14–15). After Monday's memory-led selloff dominated yesterday's issue, this cycle's podcasts did something almost counterintuitive: instead of piling onto the crash, the specialists spent their time re-underwriting the size of the build, and the numbers got bigger, not smaller. There's also a genuinely useful correction to a scary bear stat, and a fresh look at where the money still has to come from. And it lands on the morning TSMC reports, the first hard data point everyone's been waiting for. Ranked by where the dollars and the risk actually sit: 1. The spending forecast leapt again, and the whole economy is now riding on it. Facts vs Feelings, Jul 15, Ryan Detrick and Sonu Varghese, the market strategists at Carson Group, laying out their midyear outlook. This is the single most important thing in the sweep, and it cuts against the mood. Their chart tells the story in one line: at the start of 2026 they penciled in the five hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle) spending about $470 billion this year on capex, capital expenditure, meaning the money going into buildings, data centers, chips and power. Two earnings seasons later, the companies have told us the real number is $740 billion for 2026 alone. That moved the figure from 1.7% to 2.5% of U.S. GDP, and, as Varghese put it, "we're not talking about the GDP of the Maldives," we're talking about the $31 trillion U.S. economy. For 2027, the estimate has gone from $530 billion at the start of the year to almost $900 billion. And Detrick was blunt that this keeps happening: "In a month, we're going to update it, and I'm going to say it's 850… They keep going higher." Here's why a PM should care more about this than about Monday's red screen. Carson's other number is the one that turns this from a stock story into a macro story: over the last five quarters, real GDP growth ran about 2.1%, and roughly 45% of that growth, about 90 basis points a quarter, came from investment in AI-related hardware and software (and that even excludes the data-center buildings themselves). In plain terms: strip out the AI build and U.S. growth is barely above stall speed. That's the real reason this trade matters, not because the chips are cool, but because the economy now depends on the spending continuing. As they said on their own morning call with advisors: "our economy is all about AI… If AI really falls apart, we're in trouble." Two supporting data points from the same show: Micron said last week it will invest $250 billion in the U.S. by 2035 (up from about $200 billion), and Meta jumped ~15% last week after promising to keep spending, sell some assets, and make its own chip by September.

"We're not going to spend $470 billion on CapEx. We're going to spend $740 billion, in 2026 alone." Sonu Varghese, Carson Group, on Facts vs Feelings, Jul 15 2. The scary "AI demand is falling" number got corrected, by the guy who created it. Full Signal, Jul 15, Steve Martin of Silicon Data, who built the LLM token-expenditure index (and previously built a "pricing power" index at Bloomberg). If you saw headlines that "AI spending is rolling over," this is the segment to read, because it's the source clarifying what his own data actually says. Martin's index measures how much people are willing to pay for AI usage (tokens are the units AI models charge by). It's down about 20% since May, and, in his words, it "has become almost unintentionally… this AI bearish index." His point: that's the wrong read. He built it PCE-style, meaning it lets users substitute toward whatever model they actually choose. So a 20% drop mostly reflects people being smart shoppers, routing a simple request ("get me an Italian recipe") to a cheaper open-source model instead of paying top dollar for the most powerful one. His analogy: if your insurance covered everything you'd take the priciest drug every time; once you pay out of pocket, you economize. That's healthy market maturation, not a demand cliff. The corroboration is in the hardware. Martin's second data point: GPU rental rates are up 20-36% since December across the board (Blackwell B200, Hopper H100, and the older A100). The pattern is a "barbell", the newest B200 rents are up, the workhorse H100 dipped slightly at the front end, and even the roughly five-year-old A100 is flat-to-creeping-up. Translation: agentic AI (models calling other models automatically) has driven so much basic inference demand that even old chips are staying rented. His summary on where we are in the cycle: "Eventually, we're going to overbuild, but we're not there yet." Why it matters: this directly softens the demand-side bear case that dominated last week (the "token bill comes due" story). The bill is real and rising, but customers economizing is not the same as customers leaving, and it's the metric's own author saying so. 3. The real question isn't demand, it's how much more money is waiting, and why it'll get spent. The a16z Show, Jul 15, an a16z partner in conversation with a data-center intelligence analyst (the firm tracks every data center via satellite imagery, permits and regulatory filings). No named attribution beyond the show. This episode is the clearest map of the funding runway we've heard. The analyst's framing: economically-justified capex "can only grow so much," but there's an enormous pool of money that will get spent regardless. The hyperscalers can still grow capex 20-30% next year; CoreWeave and Oracle, because they tap capital markets, "can raise way more than 20 to 30%"; the giant infrastructure funds (Brookfield, Blackstone) are "turning all of their eyes" to AI infra; and the sovereign wealth funds, G42, Norway, Singapore's GIC, "have barely started touching AI." His honest caveat is the one to write down: a lot of this "is not clear from… a spreadsheet" that you should spend it, "but people will because they believe." That's the bull case and the bubble risk in the same sentence. On the competitive picture, three things stood out. First, custom in-house chips are "the biggest threat to Nvidia": Amazon is making "millions of Trainium," Google "millions of TPUs," and Google's TPUs are "100% utilized" (Amazon's Trainium isn't there yet, "but I think Amazon will figure out"). Microsoft's own silicon, bluntly, "kind of sucks." Second, he argues Google should sell its TPU chips on the open market, "not just renting, but physically", because that business alone could be worth more than the market gives it credit for. Third, and most important for the read-throughs below: the U.S. build is gated by power, not chips. "Google has a ton of TPUs sitting, waiting for data centers to be powered," as does Meta with GPUs, which is why Meta is now building data centers that are effectively tents, and why Google just bought 8% of crypto miner TeraWulf (and CoreWeave bought a ~$10 billion crypto-mining company), not for the Bitcoin, but for the power and the sites. 4. The clean framing of the bear math: $688B going in, $110B coming out. SaaStr, Jul 15, Rory O'Driscoll of Scale Venture Partners, a veteran software investor. O'Driscoll gave the tidiest version of the gap the bears keep circling. Quoting Keynes, "I'd rather be roughly right than exactly wrong", he lays it out: hyperscalers are spending $688 billion this year to make AI happen, and what's coming out the other side is about $110 billion in revenue (roughly $89 billion of it from the two big foundation-model companies, the rest rounding up). "We're spending half a trillion dollars more than we're taking in. This is a meaningful fact." His model says it takes until 2031 or 2032 before OpenAI's and Anthropic's revenue surpasses the cumulative capex, so "it's going to take probably 5 or 6 years before we're not in… invest mode." And his warning is measured, not apocalyptic: "at any point you could have a hiccup… someone wakes up and goes, 'Oh my God, what are we doing? We're spending half a trillion dollars… Maybe we should slow down.'" To close the gap, AI has to capture a big chunk of what companies currently pay knowledge workers, "well north of 25%" of software-developer dollars, i.e. "for every $200,000 software developer, they're spending $50,000 on tokens." Why it matters: this is the bear case stated fairly by someone who's still investing, the spending is real, the payback is years out, and the risk is a confidence wobble, not a proven dead end. 5. A trader's-eye view: the Goldman number, the concentration trap, and Meta's "first crack." ITPM, Jul 15, the trading desk on the Institute of Trading and Portfolio Management's "At The Desk" show (traders/PMs). Two useful things from the desk. The eye-popping stat: Goldman projects the four largest hyperscalers (Meta, Microsoft, Amazon, Alphabet) will spend a combined $5.3 trillion on capex from fiscal 2025 through 2030, up from roughly $200 billion a year across the four just two years ago to about $600 billion this year. And the risk-management lesson, which is the actionable part: being "long one energy play, one hardware name, and one industrial building data centers" is not three bets, it's "one idea… wearing three different costumes." That's exactly how Monday's selloff "double-whacked" over-concentrated books. They also flagged Meta's move to sell its excess computing power as possibly "the first really crack in the story", because the whole AI thesis rests on demand outrunning supply, and a hyperscaler turning around to resell capacity hints it may have over-ordered. Their honest hedge: "I don't think it's a full crack yet, but we have to closely monitor." They tied it to the funding theme too, Alphabet raised ~$85 billion in expanded equity, Meta is talking about a tens-of-billions share sale, and Microsoft may be next, dilution that "puts downward pressure" especially on the smaller AI names.

The debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis, the day after the market took a real swing at it, and the forecasts answered by going higher. Bull, the build is bigger than the crash, the demand is maturing (not dying), and there's a wall of money still to come. The single strongest bull fact this cycle is that the spending estimates were revised up right through the selloff: $470B → $740B for 2026, $530B → ~$900B for 2027, now 2.5% of GDP, with AI driving ~45% of U.S. growth (Facts vs Feelings, Jul 15). The scariest new demand stat, token spending down 20% since May, is rational substitution, not collapse, per the person who built the index, and GPU rental rates are still climbing even for old chips (Full Signal, Jul 15). And the funding runway is enormous and barely tapped: hyperscalers can grow capex another 20-30%, CoreWeave/Oracle far more via capital markets, Brookfield and Blackstone pivoting in, and sovereign wealth funds "have barely started" (The a16z Show, Jul 15). On the chips, Nvidia still holds pricing power (a sustained ~75% gross margin) and remains the default unless AI concentrates enough for custom silicon to win. Bear, the payback is half a decade out, the demand has no pricing power, and the money is now coming from outside. O'Driscoll's arithmetic is the anchor: $688B in, ~$110B out, and cumulative capex isn't covered by foundation-model revenue until ~2031-32, "at any point you could have a hiccup" (SaaStr, Jul 15). Token prices have fallen ~99% in three years yet bills keep rising, and open-source models mean there's "just no pricing power", margins compress as one company's spend is another's revenue in what the ITPM desk stopped just short of calling a Ponzi ("money… just kind of goes from one company to the other") (ITPM, Jul 15). The balance-sheet stage flagged last week continues: Alphabet ~$85B of equity, Meta a tens-of-billions share sale, Microsoft possibly next, the strongest companies on earth reaching outside their own cash flow to fund the build (ITPM, Jul 15). And the a16z bull himself concedes the tell: much of the incremental capex "is not clear from a spreadsheet… but people will because they believe" (The a16z Show, Jul 15).

"We're spending half a trillion dollars more than we're taking in. This is a meaningful fact." Rory O'Driscoll, Scale Venture Partners, on SaaStr, Jul 15 The synthesis. The most useful way to hold both sides at once: demand is genuinely strong and maturing (people economizing on model choice), the build is genuinely enormous and increasingly debt/equity-funded, and the payback is genuinely coming and years away. The trade doesn't break on a bad demand print, it breaks on a confidence wobble, because the money increasingly relies on belief plus outside financing rather than internal cash flow. That's why the sell signals below are mostly about funding and guidance, not about token counts. Sell signals to watch: a hyperscaler guiding flat or slowing forward capex on the late-July calls (still the trigger); the capex-to-revenue gap failing to narrow as results land; the funding mix tilting further to bonds and equity raises (widening AI-bond spreads, more dilution); Meta's "sell the excess compute" move spreading to a second hyperscaler (the over-ordering tell); Silicon Data's token index falling for a reason other than substitution (i.e., outright usage decline); and TSMC today (Jul 16), the first hard number on whether chip demand is still accelerating.

Stocks in play

NVDA. Bull: still the default winner and still holds pricing power, a sustained ~75% gross margin is the market's proof it's not commoditized, and Nvidia stays the most valuable company "for a long period" if AI stays concentrated in a few big customers. Bear: the sharpest threat is its own customers' custom chips, Amazon making "millions of Trainium," Google "millions of TPUs" that are already "100% utilized", plus the risk that cheap open-source models disperse demand away from its flagship parts. Next: TSMC's print today (Jul 16) is the immediate read-through; Nvidia's own results in August. (The a16z Show, Jul 15; Full Signal, Jul 15) AVGO. Bull: Broadcom still owns "the vast majority of share in ASICs" (the custom chips hyperscalers use to build Nvidia alternatives), and that market is growing as Amazon, Google and Meta ramp their in-house silicon. Bear: it's no longer the only game, MediaTek has become "a fierce competitor" picking off parts of the ASIC market, and the whole custom-silicon trade lives or dies on whether AI demand stays concentrated enough to justify bespoke chips. Next: custom-ASIC design-win cadence; watch whether MediaTek pursues its own U.S. listing. (The Circuit, Jul 14; The a16z Show, Jul 15) AMD. Bull: structurally, the industry still wants a credible second source to Nvidia, and AMD remains it. Bear: it barely registered in this cycle's podcasts, the custom-silicon conversation centered on the hyperscalers' own chips and Broadcom/MediaTek, not AMD. Next: mostly quiet on the podcasts this cycle; watch for the next AMD AI event and any MI-series design-win news to reset the narrative. (No fresh AMD-specific signal in the Jul 14-15 sweep.) MSFT. Bull: it's one of the five anchoring the $740B build, and, notably, it's the hyperscaler voicing the right caution, wary of building its whole stack on a single closed-source model it doesn't control. Bear: its own custom silicon "kind of sucks" versus Amazon's and Google's, leaving it more dependent on merchant chips; and it may need to raise outside capital if capex keeps climbing. Next: FY26 Q4 capex commentary in late July, still the single guide that could move the whole market if it comes in flat. (The a16z Show, Jul 15; Full Signal, Jul 15) GOOGL. Bull: the standout structural position, its TPUs are "100% utilized" and the only real alternative to Nvidia; the argument that Google should sell TPUs on the open market (not just rent them) implies a business the market isn't fully pricing. Its $80B capex is largely self-funded relative to peers. Bear: it's now funding growth with outside capital too ($85B raised in expanded equity), and it just bought 8% of crypto miner TeraWulf, a sign even Google is scrambling for power and sites. Next: late-July capex guide; any concrete move toward selling TPUs externally. (The a16z Show, Jul 15; ITPM, Jul 15) AMZN. Bull: it's building "millions of Trainium" chips, a real second path to Nvidia, and the a16z view is that "Amazon will figure out" how to get them fully utilized. Bear: it's in the same funding squeeze as the group and its build is gated by power like everyone else's; the payback math (O'Driscoll's $688B-in/$110B-out) applies squarely to its spend. Next: late-July earnings and AWS capex guide. (The a16z Show, Jul 15; SaaStr, Jul 15) META. Bull: it jumped ~15% last week after committing to keep spending, sell some assets, and ship its own chip by September; it's securing power creatively (building "tent" data centers to get GPUs online). Bear: the mixed signal of the cycle, its move to sell excess computing power is being read as a possible "first crack" (a hyperscaler with capacity to resell may have over-ordered), and it's floating a tens-of-billions share sale to keep funding the build. Next: late-July earnings; watch the capex guide and whether the "sell excess compute" plan is confirmed. (Facts vs Feelings, Jul 15; ITPM, Jul 15)

Read-throughs

  • Power / grid, still the hard physical wall, and now the gating factor by name. The single most important line this cycle came from a16z: the U.S. build is limited by "power, not chips", "Google has a ton of TPUs sitting, waiting for data centers to be powered," Meta likewise, which is why Meta is building tent data centers, Google bought 8% of miner TeraWulf, and CoreWeave paid ~$10B for a crypto miner, all for the power and sites, not the coins (The a16z Show, Jul 15). On the Texas side, ERCOT's new "batch zero" process is the successor to the SB6 deadline we've tracked: ERCOT estimates ~35 GW of "firm" new large loads plus ~65 GW of "studied and allocated", a top end of 110 GW of new demand over five years, roughly doubling the current ~86 GW system peak, with transmission the binding constraint (much of the grid dates to the 1960s), and a chicken-and-egg qualification hurdle (projects need a general contractor, a tenant, and zoning before they clear) (Energy Capital, Jul 15). Most actionable: stay long the speed-to-power and transmission enablers; the constraint is real, quantified, and worsening. Direct Vertiv (VRT) and Eaton (ETN) commentary was absent this cycle, a coverage gap, not a negative signal.
  • Memory, the epicenter, and the debate has moved to valuation, not scarcity. The Circuit's specialists framed it cleanly: SK Hynix raised $26.5 billion in its first U.S. listing (heavily oversubscribed; the stock has been "up, then down… up again"), a deal that "there is zero chance… happens if it wasn't for this memory thing." Its chairman told the roadshow the boom won't end "until 2030," and SK Hynix plans to spend ~$100B on capex in coming years. The real debate now isn't whether memory is tight (consensus: through 2028 and "very possibly… into 2030"), it's whether the earnings are durable: "they don't believe the E of the PE is sustainable long-term," which is why Michael Burry's "I will not invest in cyclicals at [high] PEs" jab landed. Watch two spin-offs/listings: SK Hynix's NAND unit Solidigm (hiring an IR head, likely to be spun out) and MediaTek as the next possible U.S. listing (The Circuit, Jul 14). Micron's $250B U.S. commitment by 2035 reinforces the supply-side seriousness (Facts vs Feelings, Jul 15). Most actionable: respect the structural tightness but treat high multiples as a warning, not comfort, the market is telling you it doubts the durability.
  • Networking / optics (Marvell, Astera Labs, Credo, Coherent, Lumentum, Fabrinet). Quiet on the podcasts this cycle, no dedicated optics or networking-silicon commentary surfaced in the Jul 14-15 window. The only adjacent read is that MediaTek competes with Broadcom and Marvell in ASICs (The Circuit, Jul 14). Carry the positions; no fresh catalyst.
  • Utilities / nuclear (Vistra, Constellation, Talen). No name-specific commentary this cycle. The generic read from the Texas discussion is that "clean firm power", nuclear, storage, geothermal, low-carbon gas, is where procurement is accelerating, and large buyers have contracted 143 GW of clean capacity since 2014 (a third of it in Texas) (Energy Capital, Jul 15). Quiet on the specific nuclear names.
  • Custom silicon / ASICs, the swing factor for the whole chip complex. The cleanest through-line across two shows: custom chips are "the biggest threat to Nvidia," Broadcom still owns most of the ASIC share but MediaTek is encroaching, and the entire trade hinges on whether AI demand stays concentrated (good for bespoke silicon) or disperses across cheap open-source models (good for merchant chips and commoditized inference) (The a16z Show, Jul 15; The Circuit, Jul 14).

What changed vs last issue

Yesterday's issue (Jul 15, "Memory cracked first, is the AI trade next?") was about the fear finally becoming price action: SK Hynix's record one-day drop, the Korea rout spilling into U.S. chips, and both bulls (Gavin Baker's "no dark GPUs") and bears (Fred Hickey's depreciation case, the AI-bond wave) sharpening their arguments. Today, the podcasts pivoted from the crash back to the size of the build, and the numbers went the bulls' way even as the mood stayed sour:

  • The forecasts got revised up, not down. This is the real delta. Carson's numbers moved from $470B → $740B (2026) and $530B → ~$900B (2027), now 2.5% of GDP, with AI driving ~45% of U.S. growth. A week of selloff talk did not dent the spending trajectory, if anything, it climbed. The macro-dependence framing ("if AI falls apart, we're in trouble") is the new lens.
  • The scariest fresh bear stat got neutralized at the source. Last week the demand-side worry was "the token bill comes due." Today the creator of the very token index the bears cite (Silicon Data's Steve Martin) said the 20%-since-May drop is rational substitution to cheaper models, not falling demand, and GPU rentals are still rising. Conviction that demand is cracking should tick down on this.
  • Meta became the ambiguous name. Yesterday Meta was the memory/power story (the ~5GW Entergy shell). Today it's both bullish (up ~15%, committing to spend + own chip by September) and bearish (planning to sell excess compute, which the ITPM desk called a possible "first crack," plus a tens-of-billions share sale). Genuinely two-handed, worth watching most closely into late-July earnings.
  • The funding/balance-sheet story continued to corroborate. Last week it was the bond wave (~$75B from Nvidia, SpaceX, Amazon). Today it's the equity side: Alphabet ~$85B raised, Meta a large share sale floated, Microsoft possibly next. The "cash printers now reaching for outside money" theme is holding.
  • The power constraint got named as the gating factor. Last week: turbine lead times and diesel gensets. Today: chips literally "sitting, waiting for data centers to be powered," hyperscalers buying crypto miners for their power, and ERCOT quantifying "batch zero" at a 110 GW top end (doubling the Texas grid in five years) with transmission the bottleneck. The successor to the Jul 15 SB6 deadline is now the batch-zero qualification process.
  • Deltas on the big numbers: 2026 hyperscaler capex ~$740B including Oracle (Carson) / $688B for hyperscalers alone (SaaStr), consistent with last week's ~$750B. 2027 now ~$900B–$1T (up from $530B at the start of the year). New this cycle: Goldman's $5.3T four-hyperscaler capex for FY2025-30; the token-spend index down ~20% since May (with the "substitution, not collapse" caveat); Micron's $250B U.S. commitment by 2035 (up from ~$200B); SK Hynix's $26.5B U.S. raise. No hyperscaler has guided capex flat, Jason Ware's trigger from last week is still unpulled.

Next prints to watch: TSMC today (Jul 16), the first hard read on whether chip demand is still accelerating, then the late-July hyperscaler earnings (MSFT, AMZN, META, GOOGL), where the one question that matters is still whether any of them dares to guide capex flat.