What this covers

Matt talks with Todd about Intel, treating it less as a stock call than as a question of how to watch a slow semiconductor turnaround. Todd begins with how the company got here. He says it missed mobile because it would not dilute PC margins by taking lower-margin iPhone volume from Apple, a contested claim. It fell behind TSMC by stretching older lithography instead of adopting EUV early, and it spent badly on Altera, Mobileye, and McAfee. Against that he sets Pat Gelsinger's plan to split design from a foundry business, on a five-year timeline that Wall Street finds uncomfortable.

The middle of the conversation is operational. Todd describes five process nodes in four years, with Intel 18A bundling four technology shifts, including gate-all-around transistors and backside power. He explains why porting customer designs into Intel's fabs is hard, why a trained workforce is both a moat and a cost, and why roughly $125 billion of capital spending will weigh on margins. He also covers why an EDA ecosystem matters, with Lip-Bu Tan's board seat as a positive, and why a Western foundry works as geopolitical insurance against Taiwan and Korea concentration.

The last stretch turns to AI. Todd discusses Gaudi3, the software stack as the real source of NVIDIA's compounding, and the need for a Moore's law of training. His most contested forecasts are that AI is a megacycle funded by hyperscalers, that monetization must show up by mid-2025, and that holders should expect a deep drawdown in 2025 to 2026. He names AMD share gains and a macro hard landing as the main risks. Throughout he stresses temperament over analysis, and Investor Overextrapolation as the trap in semis.

Sharpest takeaway

Todd Ahlsten argues Intel's recovery depends on regaining process-technology leadership and building a credible foundry business, while the AI semiconductor cycle is a larger, less cyclical megacycle. He concludes investors should track execution signposts rather than the consensus negativity, and expects a sharp drawdown before the cycle matures.

  • Intel's decline stems from missing mobile, falling behind on EUV and 10nm-to-7nm process technology, and poor capital allocation, so only process leadership can restore its moat
  • Intel's 5-nodes-in-4-years roadmap, foundry customer wins, Gaudi3 GPU orders, and EDA ecosystem partnerships are the observable signposts that will confirm or refute the turnaround
  • AI spending by self-funded hyperscalers creates a larger, durable megacycle, but investors should expect a 50 to 60 percent drawdown along the way

The argument · threads11 threads · 82 claims
0.75

Cyclical recovery and secular rerating can drive Intel's stock above very low consensus expectations.

8 pointscentrality 5/5
  • Consensus expectations for Intel are very low, so the bar for a positive surprise is only about two feet high, which Todd views as a low hurdle.

    “I think it's incredibly negative... The consensus is almost like jumping over a two-foot bar.”

  • Semiconductors combine cyclicality with a long-term secular growth trend, and the best returns come when a cyclical recovery triggers a secular rerating of the stock because investors overextrapolate negatives at bottoms.

    “when you have cyclicality around a secular trend, it's an incredible place to invest when you can get that cyclical recovery and then a secular rerate because bottoms, people always overextrapolate negatives”

  • Todd estimates Intel's client business bottomed near $29 billion in 2023 and could recover to about $34 billion by 2026, with data center CPU revenue reaching about $23.4 billion by 2026, both considered conservative.

    “their business in 2023 bottoming out at $29 billion for client... recovering $34 billion by 2026”

  • Intel's first benchmark is a return to a 25 percent operating margin by 2026, which Todd estimates implies roughly $4 of earnings power, a multiple of 18 to 19 times, and a stock price in the mid-70s.

    “our goal here by 2026 is for Intel to get back to 25% operating margin, roughly $4 in earnings power. You put an 18, 19 multiple on that. Stocks, mid-70s”

  • The current semi cycle is a cycle within a cycle because memory, GPU, logic, and PC cycles now overlap in different ways than before, and AI spending is crowding out budgets for mobile, PC, and traditional server spending.

    “this AI wave is sucking the oxygen out of some of these budgets”

  • This AI cycle is a megacycle with more durable upside than prior cycles because hyperscalers such as Google, Microsoft, Amazon, and Meta are self-funding with enormous balance sheets, unlike the debt-financed dot-com boom with unproven business models.

    “the balance sheets of the investors... are self-funding, unlike, say, the dot-com bubble where you are raising a lot of debt”

  • PC units fell from over 300 million at the pandemic peak to about 250 million, and a catch-up plus AI-enabled PC refresh could push volumes back above 300 million in 2024 and 2025, after which unit growth would be low single digits.

    “Maybe we get north of 300 million again on PC. That's a nice little recovery.”

  • Semis could keep growing and beat estimates even if the economy has a couple of quarters of negative GDP growth this year, because AI investment and the pandemic-trough recovery in PCs and servers offset the macro drag.

    “you could see potentially even a couple of quarters of negative GDP growth this year, and semis are doing fantastic”

0.69

Intel's turnaround rests on process technology leadership, subsidies, and Gelsinger's execution.

11 pointscentrality 4/5
  • The U.S. CHIPS Act's roughly $53 billion of funding offers subsidies around 25 percent of U.S. fab costs, with Intel potentially receiving $8 to 12 billion, which helps narrow Intel's cost gap versus Taiwan.

    “the CHIPS Act is a huge $53 billion. Intel has... $8 billion to $12 billion in potential CHIPS Act funding”

  • Building a fab is not enough because a skilled workforce is required, and Gelsinger has said Intel must spend about $50 million on worker training, making people and training a moat and a cost.

    “they've got to spend $50 million just training workers”

  • Intel's share and technology slide started around 2016, when it began falling behind on process technology, and the share trends in 2023 were especially poor, with the first half of 2024 expected to remain tough before improving.

    “going back to even '16 is when things started, it's hard to pinpoint a date”

  • Intel's market cap grew from about $4 billion when Andy Grove joined to a peak above $200 billion, illustrating the scale of the company's former dominance before the decline.

    “market cap of Intel was around $4 billion when he joined and it peaked at over $200 billion”

  • Pat Gelsinger has the genetics of the old Intel, centered on process, technology, leadership and pushing the laws of physics, which gives a real chance that his turnaround becomes one of the great American turnarounds.

    “he really has the genetics of the old Intel, which is about process, technology, leadership, pushing the laws of physics”

  • If Gelsinger executes, Intel could become one of the Magnificent Seven stocks within three to five years, though Todd says this is a roadmap of possibility rather than a sure outcome.

    “could Intel be one of the magnificent seven in three to five years? I'm not here to say that's a slam dunk”

  • The core bet on Intel is process technology leadership, since that is where the fixed cost is loaded, and if Intel maintains CPU share and gains even small foundry or GPU revenue, the investment is acceptable.

    “the bet around to us is it's process technology leadership because that's where they're loading the fixed cost”

  • Gelsinger is quadrupling down on the foundry bet, committing $125 billion of capital, and if any CEO can execute this plan, Todd believes it is probably Gelsinger.

    “if there's one CEO in the world that could do it, it's probably Pat”

  • Intel is a contrarian investment positioned in front of an exponential semiconductor trend, because the market prices low expectations and the company's future depends on process turnaround.

    “Intel is a contrarian investment in front of an exponential trend”

  • Semiconductor turnarounds take a long time, especially in device manufacturing, because process technology is hard and setbacks take months to fix, unlike software bugs that can be patched quickly, so investors must be patient and track signposts.

    “semi turnarounds take a long time, and that's the opportunity and the risk”

  • Backside power delivery (PowerVia) improves power efficiency and performance per square inch of silicon, which would show up as better battery life, speed, latency and AI capability, if Intel executes it.

    “backside power, where they get better power efficiency, better performance per square inch of silicon”

0.61

Contrarian temperament and low expectations favor patient positioning in semiconductor stocks.

8 pointscentrality 4/5
  • Parnassus owns NVIDIA since 2018, Intel and AMD, hedging its semiconductor exposure, and sees positive scenarios where all three win as the industry grows toward over $1 trillion by decade end.

    “we've owned NVIDIA since 2018 and remain positive on it. We also own AMD”

  • Investors who buy semiconductor leaders and hold them through cycles can do well because the cycles are long and the drawdowns are recoverable, but they must tolerate the volatility.

    “if you want to buy them, go to Maui and drink beers for seven years, and you're probably going to do okay”

  • Intel's long-term goal under David Zinsner is 60 percent gross margin and 40 percent operating margin, which with revenue above $75 billion, and possibly $100 billion by decade end, would mark a major earnings recovery.

    “the long-term goal with David Zinsner is 60% gross margin, 40% operating margin”

  • Consensus on Intel is incredibly negative because it missed mobile, fell behind TSMC at 10nm to 7nm, adopted EUV late, and faces AMD share gains, leaving declining monopolies, lower return on capital and a tough competitive battle, so the bar for expectations is low.

    “The consensus is almost like jumping over a two-foot bar. I don't think it's very high.”

  • The GPU market is more concentrated than the CPU market because a few large AI buyers such as Google, Microsoft, and Meta account for the bulk of demand, making it winner-take-all, while CPU demand is more fragmented across many buyers.

    “there's an incredible concentration winner takes all in that versus CPUs, I think, will be a bit more fragmented”

  • EDA companies such as Cadence and Synopsys are less cyclical than chipmakers because they earn from invoicing and price increases rather than unit cycles, so they should grow largely regardless of the semi cycle.

    “for Cadence and Synopsis, you're just invoicing people, raising prices, bringing out new products. So there's not as much of a cycle there”

  • Investors often believe they missed a semi trade after a 20 to 30 percent move, but because cycles last years, late entry is usually still profitable, though investors must recognize when the cycle is late.

    “don't think you missed it just because you missed the first 20% to 30% of a semi trade in a cycle”

  • Investors overextrapolate in semis, so one should be negative when everyone is positive and positive when everyone hates a stock, and temperament matters more than IQ or research depth.

    “When everybody on the earth is positive, it's time to get negative. When everybody hates it, it's time to get positive”

0.59

Intel's foundry turnaround is an aggressive, high-stakes execution test with margin and cost risk.

7 pointscentrality 4/5
  • Gelsinger is splitting Intel into a design business for CPUs in PCs and servers and a foundry business for manufacturing, aiming to make Intel an alternative manufacturer to Taiwan and Korea.

    “splitting the company in two. One side is going to be the design... Then he's splitting the foundry side”

  • Investors should watch Intel 3 and Sierra Forest and Granite Rapids launches, foundry customer deposits and announcements, Gaudi3 GPU orders, and EDA partnership announcements as the key signposts for whether the turnaround is working.

    “So look for either more customer deposits or customer announcements for Intel foundry solutions.”

  • Intel is attempting five process nodes in four years, starting with Intel 7 and continuing through Intel 4, Intel 3, Intel 20A, and Intel 18A, which Todd views as aggressive architecture adoption that is the first test of the turnaround.

    “they're going to go through five nodes in four years, which is really aggressive architecture adoption”

  • Intel 18A combines four simultaneous technology shifts: EUV, gate-all-around transistors, PowerVia backside power delivery, and new chip-on-wafer-on-silicon packaging, which is why Todd calls it four waves in one.

    “Intel 18A, which is really four waves in one”

  • Intel's commitment to designing and manufacturing chips in-house means any design misstep also hurts manufacturing efficiency, so an error in architecture leverage compounds into a cost disadvantage.

    “Intel is a double bet on we're going to design it here and build it here. So if you make a small error on designing it... you can't build it as effectively.”

  • Intel Foundry Solutions will drag on margins for a long time because Intel is putting roughly $125 billion of capital into the ground over five years and depreciation already exceeds $10 billion per year on depressed revenue.

    “We're running depreciation over $10 billion a year... The foundry business will be diluted for quite some time. You've got a huge fixed cost.”

  • Intel cannot invest $125 billion at a 30 percent cost disadvantage to Taiwan and still earn a return, so closing the cost gap through subsidies, process catch-up, and technology is a precondition for the foundry business.

    “Intel can't plow $125 billion of capital at a 30% cost disadvantage to Taiwan. There's no business there.”

0.57

AI cycle timing and customer monetization determine whether overbuilding causes drawdowns.

12 pointscentrality 3/5
  • Because AI investment is so large and the buyers are well financed, the semiconductor industry may be less cyclical in the short term, though whether that holds long term depends on how AI models are monetized.

    “in the short term, it's going to make the industry less cyclical. We're going to have to see how these models get monetized”

  • The AI accelerator market could reach about $400 billion by 2027, which, per New Street Research's analysis, implies roughly $1.1 trillion of data center spending, a bullish estimate that is itself a sanity-check concern.

    “if you look at a $400 billion accelerator market by 2027, that's 1.1 trillion of data center spending.”

  • Investors holding NVIDIA and Intel for seven years should expect a 50 to 60 percent drawdown at some point, possibly 18 to 30 percent, likely in the 2025 to 2026 window, because the market may overbuild before monetization is proven.

    “there's highly likely that 50% to 60% drawdown at some point”

  • Semi stocks bottomed around October 2022, and recent semi upturns have lasted at least two and a half years, so the current cycle may run longer than typical, with the next 12 months looking strong.

    “a lot of these upturns in the last decade have been at least two and a half years. Things started to... bottom out October 2022.”

  • The key unanswered question for the AI megacycle is whether AI business models such as copilots, digital ads, medical research, and precision agriculture can generate enough revenue to justify hundreds of billions of infrastructure spending.

    “There's going to have to be a lot of business models that kick in.”

  • AI GPU spending is crowding out other data center and PC budgets, so the AI wave could top out around mid-2025, after which mobile, PC and traditional server cycles could grow as digestion occurs.

    “this AI wave is sucking the oxygen out of some of these budgets”

  • The critical question for the AI cycle is whether hyperscaler and enterprise customers can monetize AI, such as Intuit with QuickBooks or John Deere precision agriculture, because a train-now, pay-later mismatch is the biggest risk.

    “There's going to have to be a lot of business models that kick in.”

  • The semiconductor cycle may be longer than usual, with the next 12 months looking constructive, because memory is emerging from its downturn, logic and GPU are set for recovery, and AI requires compute across memory, NAND, DRAM and edge.

    “a lot of these upturns in the last decade have been at least two and a half years. Things started to... bottom out October 2022.”

  • Semiconductor stocks typically peak before the underlying industry cycle peaks, so waiting for operating peak data can be too late for investors.

    “obviously, the stocks will peak before the cycles peak”

  • The AI ecosystem cannot be only GPUs; it requires compute at the edge, in memory, NAND, DRAM and PCs, which broadens the demand base for CPU and memory suppliers.

    “They can't do it alone. There has to be compute on the edge, in memory, NAND, DRAM, PCs”

  • The central question is whether AI businesses monetize the training build-out by mid-2025, and if not, the AI wave could top out then, with other markets like mobile and PC growing as budgets return.

    “you could have a case where AI, for the short term, tops out mid-2025.”

  • The semiconductor industry, currently about $500 billion, could grow well above $1 trillion by the end of the decade, implying a lot of silicon and people needed to build it.

    “semi industry going to well over $1 trillion by the end of the decade”

0.47

Mobileye stake, Western foundry, and asset monetizations could unlock value the market is not paying for.

8 pointscentrality 3/5
  • Geopolitical risk around Taiwan and Korea makes Intel's push to be a Western foundry valuable as a supply-chain insurance policy, because it adds manufacturing capacity in the U.S., Germany, and Israel outside the Taiwan-Korea concentration.

    “Geopolitics are super intense here, which I feel break positive for Intel... it's almost like you call up Allstate or GEICO”

  • Intel owns over 80 percent of Mobileye, which Todd values at about $25 to 30 billion, and he considers it a call option on autonomous driving and China EV growth that is not fully priced into Intel's shares.

    “they own roughly over 80% of that. On paper, it's a $25 billion, $30 billion cap company”

  • Todd's bull scenario has Intel's GPU business reaching $2 to 3 billion in revenue within three years and Intel Foundry Solutions reaching $3 billion in three years, with foundry potentially exceeding $10 billion by the end of the decade.

    “Could Intel Foundry Solutions be $10 billion plus?”

  • Intel's programmable solutions business, the former Altera, and its EUV mask-writing unit, which raised outside capital at roughly a $4 billion valuation for a 20 percent stake, could be monetized or spun out to unlock value.

    “They sold 20%... that could be potentially enormous”

  • Intel can win GPU share by targeting a good-enough segment at lower price points, since not all model training requires top-tier NVIDIA hardware, though Intel must first prove performance and software competitiveness.

    “Intel could really play in that good enough market”

  • If Intel Foundry Solutions crosses $10 billion in revenue by the end of the decade, it could earn roughly 20 percent returns on capital, a figure the company has floated in analyst day materials.

    “Let's just call it 20% returns on capital, potentially.”

  • At a roughly $200 billion market capitalization, investors are not paying for Intel's GPU and foundry optionality, which is why Todd sees the investment as getting intriguing, roughly a ground-floor valuation if the plan works.

    “at the current valuation, at $200 billion, you're just not paying for that right now”

  • Todd ranks the investment opportunity as Intel first, NVIDIA and AMD second, and Cadence and Synopsys third, because Intel is the tip of the spear on the CPU, PC, and server cycle with the most turnaround leverage.

    “So I think the opportunity is highest on Intel, one, NVIDIA and AMD, two, and then Cadence and Synopsys, three”

0.43

Intel's past errors in mobile, process, capital allocation, and customer focus explain its decline.

6 pointscentrality 3/5
  • Intel missed the mobile market because it was too PC-centric and unwilling to dilute its high-margin PC cash cow by accepting lower-margin volume chip orders from Apple for iPhones.

    “they were too PC-centric. They weren't willing to dilute the cash cow, the gravy train. So they missed mobile.”

  • Intel's share of servers and PCs fell from roughly 90 percent to about 80 percent in PCs and 70 percent in servers, and Todd expects further erosion into 2024 until new products launch.

    “close to... 90% share in servers and PCs. Now that's dropped to 80% share in PCs and say 70 in servers”

  • Intel's gross margin fell 17 percent from its peak, operating margin dropped from about 30 percent to 7 percent, and R&D rose to 7.5 percent of sales, with revenue falling from a $78 billion peak to about $54 billion.

    “gross margins down 17% from the peak, R&D spending up 7.5% of sales, operating margins going from 30% to 7%”

  • Intel fell behind TSMC at the 10nm-to-7nm transition because it relied on double patterning and immersion lithography to squeeze older technology rather than adopting EUV early, which TSMC did.

    “they went into things like double patterning and immersion, different lithography that they tried to get more out of older technologies and didn't really adopt EUV.”

  • Intel was a poor capital allocator, spending roughly $15 billion on Altera, $15 billion on Mobileye, and $8 to 9 billion on McAfee, churning capital into the wrong places.

    “They spent $15 billion buying Altera, $15 billion on Mobileye. They spent $8 billion or $9 billion on McAfee”

  • Intel lost its customer-centric orientation because it is used to designing its own products, and becoming a foundry requires porting third-party designs through the EDA ecosystem, which demands a customer-centric mindset.

    “Intel is used to building their own designs. And so when you have to go fabless... It takes a lot to port those designs”

0.43

The turnaround takes years, so patience and specific signposts like Intel 3 and Gaudi3 matter.

6 pointscentrality 3/5
  • Semi turnarounds take a long time, and Gelsinger's own five-year timeline, given at his February 2021 arrival, contradicts what Wall Street wants to hear, so patience is required on signposts.

    “semi turnarounds take a long time... he said this is a five-year turnaround”

  • The main risks to Intel are continued AMD share gains, a macro hard landing that delays PC refresh and hyperscaler capex, and execution slips, with 8 or 10 positive factors versus 3 or 4 negative ones.

    “we have a hard landing on the economy... That could be tough.”

  • Porting customer designs into Intel's fabs is very hard because it requires tape-out, yield, defect, and cost management across idiosyncratic designs, so setbacks take months, unlike software patches.

    “You have a mask and EUV. This isn't something you go in and a couple guys in bunny suits... fix it. It takes six months a year.”

  • Intel 3, the second-generation EUV node, is the near-term technical signpost, and investors will know within 12 to 18 months, by end 2025, whether Intel's GPU push is working.

    “we're going to know by the end of 2025 if this is going to happen or not”

  • Intel's GPU business is still at the small-slice stage, and the real inflection would be owning customers directly, since a $3 billion slice of a $100 billion market is the first step and Intel is not yet in the final bake-off.

    “Matt, if they start owning customers, that's when you pop the champagne cork.”

  • Gaudi3, launching later in 2024, is the test of Intel's AI GPU ambitions, and the key evidence is big purchase orders rather than blog-post specs, which Todd says have not yet appeared.

    “It's one thing to talk about specs, it's one thing to deploy it. We got to see big orders.”

0.29

Stabilizing CPU share through nodes, AI PCs, and account wins is the near-term test.

6 pointscentrality 2/5
  • The first signpost to watch is progress on five nodes in four years, especially second-generation EUV products like Granite Rapids and Sierra Forest, which should stem share loss to AMD.

    “Number one here is five nodes, four years. How are we doing on that with the announcement?”

  • The next driver for Intel's CPU share is AI-enabled PCs, along with Meteor Lake and backside power, which Todd expects to stabilize share in the second half of 2025 and possibly allow share gains.

    “the first thing is let's stabilize the share in the second half of the year, 2025 with the AI-enabled PC”

  • Some Apple business, such as the Apple laptop chip business that existed through 2020, has been lost and is unlikely to come back.

    “Some unfortunately has been lost and probably isn't coming back like the Apple business.”

  • Pandemic-era PC and server distortions plus AI budget crowding out mean that current CPU demand is hard to read, so a rebound in 2024 is possible but long-term CPU market growth may be only 7-8%.

    “long term, could we see their markets grow 7% to 8%? That would set up well for Intel.”

  • Some lost share, like the Apple business, probably will not come back, so Intel's CPU recovery must come from process nodes, AI-enabled PCs, and account-by-account proof points rather than reclaiming lost customers.

    “Some unfortunately has been lost and probably isn't coming back like the Apple business.”

  • The key for the stock is stabilizing CPU share in 2H 2024 into 2025, which would allow investors to underwrite the base business, and gains would come account-by-account rather than from one customer.

    “the key thing here for the stock is can they stabilize the share, then we can start underwriting to the base business”

0.26

Winning against NVIDIA requires foundry ecosystem, customer commitments, and software.

5 pointscentrality 2/5
  • Intel's foundry needs an EDA design ecosystem and software compatibility to attract customers, so new EDA partnerships are a signpost, and Lip-Bu Tan's board addition, with Cadence experience, is a positive sign.

    “building an ecosystem is so important to put it in a foundry”

  • The CHIPS Act offers roughly a 25% subsidy for building in the US, so the $53 billion program implies a total project scale of four to five times that amount when grossed up.

    “you're getting a 25% subsidy. ... So you gross up that by four, five times.”

  • Some customers may want a non-NVIDIA alternative for patriotic or supply reasons, but that alone is not enough and Intel must become competitive on performance and software.

    “I think there's a certain -- even a patriotic, I'm rooting for Intel's GPU and the foundry”

  • Intel has missed the AI GPU wave so far, and despite Gelsinger previously working on GPUs years ago, it has not yet built a software stack good enough to win large orders.

    “it just hasn't been there yet with the software stack”

  • The second signpost is foundry business model evidence: customer down payments or deposits and named customer announcements for 18A, since they show external demand for Intel manufacturing.

    “look for either more customer deposits or customer announcements for Intel foundry solutions”

0.24

AI training and inference demand favor GPUs and CPUs, with software stacks as the key moat.

4 pointscentrality 2/5
  • GPUs can compound like Moore's law squared because the software stack around them, as with NVIDIA's CUDA, creates economics that hardware alone does not, and Intel's Gaudi software from the 2019 Habana Labs acquisition is meant to play this role.

    “I like to think of GPUs as Moore's law squared... there's a software component to these that gets really interesting. That's what Gaudi3 is”

  • Training large AI models is too expensive, so the industry needs a Moore's law of training that reduces cost per model through combined GPU, packaging, and software improvements, enabling more complex models.

    “there's a Moore's law of training. It's too expensive to train models right now.”

  • As trained AI models move to inferencing at the edge, some inference workloads may run on CPUs such as fifth-generation Xeon, which could create a CPU demand tailwind even as GPUs dominate training.

    “Some of that inferencing math is going to be done with CPUs and maybe a fifth gen Xeon processor can do that.”

  • Benchmark claims and blog posts about Gaudi2 do not prove commercial viability; only large invoices will show that Intel's GPUs are competitive, and the software stack is still the key gap.

    “that has to turn into a large invoice. And we haven't seen that yet.”