Newsletter · · Ashutosh Agarwal

Intel's Foundry Turn Gets Real as Capex Spending Spooks the Market - Foundry & Chip Equipment Weekly - Week of July 26, 2026

Foundry & Chip Equipment Weekly for the week of July 20 to 26, 2026. Intel's contract-manufacturing turn finally put signed multi-year customers, a 2028 date for 14A, and Apple on the table, even as the market punished Alphabet, Tesla, and Intel itself for raising capital budgets. TSMC hiked capex and list prices, ASML's order book kept filling, and a Cerebras CEO named the three sold-out chokepoints in the AI supply chain.

Foundry & Chip Equipment Weekly

Week of July 26, 2026: Intel's Foundry Turn Gets Real as Capex Spending Spooks the Market

Last week the puzzle was that the two best businesses in the chip supply chain, TSMC and ASML, reported spectacular numbers and their stocks went nowhere. This week the market took that mood and hardened it into a rule: raising your spending plans is now something to be punished for. Alphabet lifted its capital budget and lost about $300 billion of market value in a day. Tesla did the same and had its worst day in more than a year. The semiconductor index fell 13% in a month.

And then Intel, of all companies, reported, raised its own spending, and for one glorious after-hours moment jumped 13%. It gave most of that back by the end of the week. But underneath the noise was the most important thing to happen in the foundry world in months: Intel's long-suffering contract-manufacturing business finally put real numbers and real dates on the table. Customers signed up. A production date got set. And the people who cover this stock for a living came away sounding, for the first time in years, cautiously convinced.

So this week is really two stories running side by side. One is a company clawing its way back. The other is a market that has suddenly decided the AI build-out costs too much. Let's get into both.

TL;DR

  • Intel's foundry finally showed its hand. Management said it has locked in 10-plus customers on long-term agreements (3-to-5-year deals on price and volume), committed to putting its most advanced manufacturing process, 14A, into high-volume production by 2028, and named Apple as a new manufacturing customer. It is targeting $15 billion of outside foundry revenue by 2030, and the CEO says he can see "billions of dollars" of advanced-packaging business building up. Intel's own data-center business grew nearly 60%.
  • But it raised spending, too: capital expenditure goes to about $20 billion this year (up from roughly $15-18 billion) and "significantly higher" next year. In this market, that is a warning label. The stock popped 13% after hours, then faded and closed the week lower.
  • TSMC keeps pressing its advantage. It raised its capital budget to north of $60 billion, is putting 70-80% of it into leading-edge process and packaging, added another $100 billion to its US plans, and, per multiple reports, will raise chip prices 5-10% in January 2027, with a 10-15% surcharge for customers who under-forecast and then scramble for more.
  • ASML's order book keeps getting better. Morningstar raised its 2030 revenue estimate for the machine-maker from $60 billion to $70 billion and lifted its fair value on the stock, citing a 30% capacity expansion for 2027 (already sold out) and another 30% penciled in for 2028.
  • The bottleneck isn't the chip, it is everything around it. The CEO of Cerebras laid out the three things the whole industry is short of right now: high-bandwidth memory, TSMC's CoWoS packaging, and 3-nanometer factory space, all sold out.
  • On the frontier: IBM's CEO said quantum is now "in the engineering realm, not the science realm," and PsiQuantum's plan to build a computer out of light gave a rare, concrete look at the strange new supply chain: liquid helium, homemade crystals, and chips baked at a GlobalFoundries plant in upstate New York.

What's New

Intel: The Quarter Where the Foundry Story Stopped Being a Promise

For years, Intel's pitch to investors about becoming a contract chip manufacturer, a "foundry," building chips for other companies the way TSMC does, has been long on vision and short on proof. This week that changed.

Start with the operators, because they said the most concrete things. On CNBC's Fast Money (July 23), CNBC's Christina Partsinevelos relayed what CFO Dave Zinsner told her directly off the earnings call: Intel has "locked in 10-plus customers on long-term agreements on both price and volume," deals that "range anywhere between 3 and 5 years." They are "raising their capital spending meaningfully next year." And, the genuinely new part, they are "committing to ramping 14A, that is their most advanced process, into high-volume production by 2028. So that's new news on the foundry side." She added that CEO Lip-Bu Tan has "a line of sight to, quote, billions of dollars" in advanced packaging, "and they're already starting to get that backlog there."

Why did that land when everything else this earnings season got sold? Partsinevelos put her finger on it: "TSMC posted its best quarter in history and fell. ASML, Texas Instruments raised outlooks and yet their stocks fell. Samsung printed record profit and yet fell. But Intel locked in these long-term agreements... Intel now has visibility into demand. They know it's coming, and that's why they're raising CapEx." In other words, Intel gave the market the one thing the blowouts did not: a signed, multi-year promise from actual customers.

Bloomberg Tech's Ed Ludlow, who spoke to both the CFO and the CEO (July 23), added the capital-spending detail: this year's budget goes "about $20 billion," up from a previous guide near $18 billion, and it would grow year on year. Crucially, Ludlow said, the money is going "on the tooling side specifically, not the space side," meaning Intel is not pouring concrete for new buildings; it is filling shells it already built with chip-making equipment. And he stressed the CEO's temperament: "Lip-Bu Tan is a very cautious guy. He does not deploy capital and improve or increase CapEx unless he's sure he's going to get a return on it."

There is one honest caveat the operators kept flagging: still no named foundry customers, beyond Fortinet (announced earlier in the week for security chips) and Apple. Ludlow explained Intel's policy, "when a customer signs up to use our latest chipmaking process, we don't announce it, we leave it to the customer," but conceded the awkward result: "none of the customers have announced it. All they'll say is that they signed loads of... LTAs in the first 6 months of this year." On The Rundown (July 24), host Zaid Admani of Public.com summed up the scoreboard: the foundry business "grew 31% to $5.8 billion," Intel "recently added Apple as a customer for some of their chip manufacturing," and management is "seeing a lot of interest for their next-generation 14A manufacturing process. And also advanced chip packaging."

Why the Stock Jumped, Then Sagged

The market's reaction was a story in itself. On Fast Money, Gene Munster of Deepwater Asset Management, an investor, not an Intel insider, captured the mixed feeling. On the plus side, Intel guided revenue up 11% for the September quarter and "there's a lot of good news around 14A, which is a driver in 2028." But "they just raised CapEx from $15 billion where the Street was at for this year to $20 billion. They said next year will be significantly higher. That has been the tagline for stocks going down." His blunt bottom line: with the stock up about 12% on the week while Google fell 8%, "it just feels like a gift."

Munster also put the valuation in perspective, and it is sobering: on 2027 estimates, "Intel trades at 61 times and you have Nvidia trading at 17 times. AMD... trades at 31 times." His point was not that Intel is doomed, it is that you are paying a very rich price for a turnaround that still has to happen. As Deepwater sees it, Intel's hottest segment, data center and AI, is "one-thirteenth the size of Nvidia's GPU business." It grew 56%. That is real, but it is growth off a very low base.

By the following day, the glow had worn off. On Closing Bell Overtime (July 24), HSBC's Global Head of Semiconductor Research, Frank Lee, framed it precisely. Intel "basically delivered more or less the same print they did last quarter, a big beat and raise, but had a very different response this time." The difference was not Intel; it was mood. "If you tie it all together, hyperscaler CapEx spending is where people have concerns."

Lee's larger observation is the one to carry forward: the easy phase of this trade is over. "From about April to the end of June, you had that last stage where stocks are going up on momentum, on the long-term [market-size] opportunity," investors chasing "wherever the price hike is going to be next," with memory the poster child. "Now the market's going to come back a little bit more to the near term of which companies can, on a company-by-company basis, deliver upside in earnings relative to where the Street is." Translation: the tide that lifted everything is going out, and from here you have to pick.

For all that, HSBC still likes Intel enough that its price target implies the stock roughly doubles. The reason is exactly the foundry news: "They've talked about high-volume production of Intel 14A in 2028. That's a very significant sign that they are seeing new customers 2 years from now, giving that conviction to start high-volume production. So the foundry now is definitely worth something." He thinks the market is still underestimating it.

The Bear Seat on Intel, Voiced and Worth Hearing

Not everyone bought it, and this is where the debate genuinely had two sides this week. On Bloomberg Tech (July 24), Antoine Chkaiban of New Street Research, neutral rating with a $115 price target, argued the good news is already in the price. Even the print's most encouraging line, foundry operating margins up 9 points to head toward break-even, does not change the long-run math: "Even if you assume in 2030 Intel is fully recovered, they're going to still have some structural cost disadvantages compared to TSMC, and gross margins are never going to reach the same level." Meanwhile Intel keeps ceding ground to AMD and Arm, and to the hyperscalers' own in-house chips. "Even the bull case doesn't make much room to the current levels the stock is trading at."

But even Chkaiban conceded the foundry evidence is turning tangible. On that $15 billion outside-revenue-by-2030 goal: "There is increasing evidence that this is actually pretty tangible." And on the capex, echoing the CEO, he noted the money is "for wafer fab equipment... tools, not space," because Intel has "a whole lot of space" in already-built shells. The open question is 2027: whether Intel can keep improving yields and costs enough to make foundry actually profitable, since its margins there are still negative.

The most quotable skeptic-turned-believer was Bernstein's Stacy Rasgon on The Exchange (July 24), who has spent a career being negative on the stock. He was withering about the old Intel, "arrogant is absolutely the right word... I used the words fat, dumb, and lazy at one point," but notably softer on the new one: "I feel better about it than I have in a long, long time, because you can see what they're doing, and they are executing it." He thinks they "got lucky, frankly, on the server upside," and that today's very tight chip supply "brings the foundry narrative back into play," because customers are giving Intel "a harder look than they ordinarily would because of the nature of the current supply situation." Bernstein is still only neutral, and Rasgon was careful: "It is still a bit of a slog. They've got a lot of work to do."

TSMC: Raising Everything, and Finally Raising Prices Out Loud

While Intel grabbed the headlines, TSMC quietly kept doing what it does, spending enormous sums and, this week, telling customers prices are going up.

On The Circuit (July 20), industry analysts Ben Bajarin and Jay Goldberg walked through a quarter that beat on revenue and profit, guided up, and, most importantly, raised the capital budget to "north of 60 billion," with "70 to 80% allocated to advanced process" and packaging. Their reframing of why TSMC has to keep building is the clearest on the tape: in the smartphone era, tiny chips packed hundreds of millions per wafer; in the AI era, chips are huge, you get far fewer per wafer, and you need extra wafers just for the packaging that stitches them to memory. The result, Bajarin said, is that you can "envision four to five additional foundries needed for TSMC to meet demand." He noted the ASML machines needed for the latest nodes are "at least two and a half times the size of prior nodes," so you literally need more buildings "to make less chips than you did in a smartphone." That physical reality is why he and others now think wafer-fab equipment could be a $300 billion market by 2030.

On price, Bajarin did the arithmetic that management will not say out loud: "If you look at their numbers, you can back into the fact that they raised prices, call it 30% roughly." What is striking is how little resentment it draws, "there's a lot of bad blood now towards Nvidia, towards Micron, on their prices. And I don't get that vibe from how people talk about TSMC." (He was honest that some insiders disagree: "I know people say, oh, it's much worse than you think. Just... everyone is afraid to say anything.")

The formal price move surfaced separately. On Mac OS Ken (July 22), host Ken Ray relayed reporting (from Nikkei Asia via Apple Insider) that TSMC will lift base chip prices 5-10% starting in January 2027, and that customers who order more than they originally forecast will face an extra 10-15% surcharge, meaning a company that misjudges demand could pay up to 25% more than today. TSMC's line: the increases are "strategic, not opportunistic," driven by the cost of equipment, materials, and building new plants. For Apple, the biggest and oldest customer, the hit is expected to land nearer the 5% end. And Network Break (July 20) captured the other headline, TSMC "chips in another $100 billion for US fabs," on a quarter where revenue hit $40.2 billion (up nearly 34%), profit rose 77%, and chips at 7 nanometers or smaller made up 77% of sales.

ASML: The Order Book That Keeps Validating the Whole Boom

The machine-maker did not report this week, but it got a fresh vote of confidence. On The Morning Filter (July 20), Morningstar's Chief US Market Strategist Dave Sekera explained why the firm hiked ASML's fair value so sharply after its recent quarter, and it was not the quarter itself. It was capacity. ASML is "increasing capacity by 30% in 2027, and... potentially another 30% in 2028." Morningstar reckons the 2027 slots are "essentially already booked," with roughly 30% of the 2028 expansion booked too. Feeding that into the model, Sekera said, "our 2030 revenue forecast was $60 billion, which was the high end of their longer-term guidance. We're now expecting that to be $70 billion," and Morningstar lifted its fair value on ASML's US-listed shares to $2,050. It did the same for TSMC, raising fair value 27% to $534 per share on a 2026 capital budget hiked $8 billion to $62 billion.

Which brings us back to the puzzle from last week, and Sekera answered it squarely: why did both stocks pull back even as he raised his targets? Because "the market's already priced in the exceptionally strong growth... these companies have already scheduled their production runs for 2027." The fight now "is becoming about trying to forecast 2028," where "there's a pretty wide range of assumptions." Layered on top is a technical overhang he pinned on Korea: the Korean market, "over 50% of [whose] market cap is just two stocks," memory giants Samsung and SK Hynix, has peaked and fallen into a bear market, down about 25% from its high. When the most crowded corner of the trade unwinds, it drags the neighbors down with it.

The Debate

The bull case is grounded, once again, in order books and hard commitments. ASML's 2027 capacity is sold out and 2028 is filling; TSMC is putting 70-80% of a $60-billion-plus budget into leading edge and packaging and raising list prices into 2027; Intel got real customers to sign 3-to-5-year deals and set a 2028 date for its most advanced node. When suppliers this far up the chain are this confident about orders two and three years out, it is the strongest available evidence that the AI build-out is not a one-year sugar rush.

The bear case, this week, was almost entirely about price and digestion, not about the businesses breaking. The clearest expression came not from a chip skeptic but from the market's own behavior: Alphabet and Tesla were punished for raising capital spending, and the read-through (voiced by Bloomberg Intelligence's Mandeep Singh on Bloomberg's instant-reaction panel) is that a trillion dollars of hyperscaler capital spending in 2027 is now a number that makes investors nervous rather than excited, even as it is unambiguously good for the chip-equipment names those dollars flow to. Frank Lee's framing is the one to hold onto: the momentum phase is over, and the market is shifting to rewarding near-term earnings delivery company by company. Morningstar's Sekera named the specific fear behind the mood, that AI is "increasingly hazy" on a medium-term view, with users burning more and more tokens for the same output, cheaper Chinese models nipping at margins, and a real question of whether there is "already enough compute capacity out there for what's currently needed."

Note the shape of it honestly: no one on the podcasts this week argued that TSMC's or ASML's demand is fake or their moats are cracking. The Intel debate had two genuine sides: New Street's "even the bull case is priced in" against HSBC's "the foundry is finally worth something." But on the two crown jewels, the disagreement was purely about how much good news the price already reflects. That is a debate about valuation, not about the story.

Read-Throughs

The bottleneck has moved, and it is instructive to hear it from an operator. On The MAD Podcast (July 23), Cerebras CEO Andrew Feldman laid out the "three major bottlenecks right now," and every one is a place to look for winners. First, high-bandwidth memory (the stacked DRAM that feeds AI chips), made by only three companies in the world and "sold out," where the only way to add supply is to build new factories. Second, TSMC's CoWoS advanced packaging, the silicon "motherboard" that Nvidia and AMD use to bolt their chips to that memory, "sold out... highly constrained." Third, 3-nanometer factory space at TSMC, where "the bulk of the GPU world" is crammed in, causing "tremendous congestion." Feldman's own wafer-scale approach sidesteps all three (his chips run at 5nm and skip HBM), which is his pitch, but his map of the shortages is the read-through: memory makers, the packaging tool-chain, and leading-edge capacity are where the scarcity value sits. And his bubble check is worth keeping: unlike the 1990s telecom build-out, where "we built out fiber years before it was going to be used," today "we're all trying to catch up... for demand that's already here today."

Wafer-fab equipment is the quiet winner of the capex revolt. The detail that both the Intel CEO and two separate analysts stressed, that Intel's rising budget is going to "tools, not space," is a direct positive for Applied Materials, Lam Research, KLA, and Tokyo Electron. Filling already-built shells with equipment is the fastest, most equipment-intensive way to add capacity there is. (Worth flagging for continuity: those pure-play tool names went undiscussed on the podcasts as standalone stories this week, they report later in the earnings calendar, so this is a read-through, not a fresh operator soundbite.)

Fabless customers face the price squeeze next. TSMC's January 2027 hike, and its warning that overseas fabs run at lower margins, is really a preview of who pays for pricier chips, and the surcharge structure (up to 10-15% extra for under-forecasting) rewards the customers with the best demand visibility and punishes the rest. Apple, as the oldest and largest client, looks set to be shielded at the low end; smaller or lumpier buyers less so.

China-domestic toolmakers and Samsung's foundry drew no coverage again this week: no discussion of SMEE, Naura, AMEC, or Piotech, and Samsung surfaced only as a memory maker caught in the Korean-market unwind, not as a contract manufacturer. Both remain live risks to watch; neither moved this week.

Quantum hardware, the long-dated frontier, had its richest week in a while, and it is increasingly a supply-chain story. The standout operator voice was IBM CEO Arvind Krishna on Bloomberg Talks (July 23): "When things are two years away, I wouldn't call them a moonshot. Quantum is now in the engineering realm as opposed to the science realm." He reaffirmed IBM's target to "deliver a machine... a large-scale, fault-tolerant quantum computer in the next two years," by 2029, and sized the prize at "about a trillion dollars" of value created by the end of the 2030s. To get there, IBM said that same morning it had bought HRL Laboratories from GM and Boeing, to add talent in "materials, spintronics, quantum sensing and other sensor technologies."

The most vivid look at what building one of these things actually requires came from MIT Technology Review's narrated feature on PsiQuantum (July 22), the company trying to build a quantum computer out of light. The physical picture is the read-through: about 100 stainless-steel cabinets cooled with liquid helium to 2 kelvin, roughly -456°F, housing chips through which single photons fly. The company makes its own light-routing crystal (barium titanate) in-house, then ships the wafers "to the chipmaker GlobalFoundries in Malta, New York, where PsiQuantum's chips are made." The bet, tellingly, is that this exotic chain "looks like a souped-up... version of the existing supply chain for silicon photonic chips" already used in data centers, so it can lean on tools and infrastructure that exist. One neat cost wrinkle for the photonic approach: only the single-photon detectors need to be ultra-cold, and PsiQuantum said it will spend part of its $100 million CHIPS Act award on exactly those. The company aims to have its Australian site "operational," cooling in place and hardware-ready, by late 2027; the DARPA program chief now thinks someone builds a genuinely useful quantum machine "by 2033."

And the government money keeps organizing a domestic quantum supply chain. Crain's Daily Gist (July 21) reported that a tri-state partnership won a $30 million federal grant toward the national quantum supply chain, part of $169 million spread across six regional tech hubs. Illinois is putting $11 million into a factory for quantum-component manufacturing at its Quantum and Microelectronics Park (where PsiQuantum, IBM, and IonQ are all involved), atop a $500 million state commitment that will include "one of the world's largest cryogenic facilities." The investable thread here is not the qubit, it is the cryogenics, control electronics, specialty crystals, lasers, and packaging that any of these machines will need, whoever wins.

What Changed From Prior Weeks

For most of July this section led with a lament: no fresh operator voices from the foundry ecosystem. That is now emphatically over. Two weeks ago TSMC and ASML reported; last week Intel reported, and this week we got the foundry-specific substance from Intel that had been missing for months: signed multi-year customer agreements, a 2028 high-volume date for 14A, a named marquee customer in Apple, and a hard $15-billion-by-2030 revenue target that even a skeptical analyst now calls "tangible."

The market backdrop also flipped. The memory euphoria that defined late June and early July has curdled into a broader anxiety about the cost of the whole build-out: capex is suddenly a liability, and the Korean memory names that led the boom are now leading a pullback. The supply story, meanwhile, only got more concrete: where we used to talk about HBM in the abstract, this week an operator ticked off exactly which three chokepoints, memory, CoWoS packaging, and 3nm space, are sold out.

Still quiet, for continuity: Samsung's foundry (its 2nm yields and Taylor, Texas plant went unmentioned again), the pure-play tool makers (Applied Materials, Lam, KLA, Tokyo Electron, with no standalone coverage; they report later), and the China-domestic equipment names, which drew nothing this week.