Week 10: Intel (EE 292P — Atoms, Bits, and the National Interest)

Stanford EE professor, TSMC vet, and Hanover venture partner Dr. Ali Keshavarzi and Hanover founder Joe Malchow created EE 292P and are teaching it at Stanford winter quarter 2026. Called “Atoms, Bits, and the National Interest,” or ABNI, we are attempting to trace the build-up of national power resulting from the growth of the semiconductor industry — from transistor physics through to great power competition.

We’re bringing in leaders from key companies in tooling, foundry, design, logic, memory, power electronics, AI applications, and other fields including economics and law. This week, we held our session on economics and understanding labor and total factor economic productivity.

March 10, 2026 · Stanford University
Guest: Prof. Robert Burgelman, Stanford GSB
Instructors: Ali Keshavarzi, Joe Malchow · Also present: Prof. Tom Lee, Stanford EE

Lane 200–205
Stanford

Intel’s trajectory is the defining case study in semiconductor strategic leadership — and, as Burgelman said at the outset with no attempt at drama, “a fairly sad story.” At its peak in August 2000, Intel was valued at $509 billion — roughly $1 trillion in 2026 dollars — and briefly the most valuable company in the world. When Pat Gelsinger was removed in late 2024, the market cap stood at $100 billion. One-fifth of the peak. The session traced, in forensic detail, how that happened — and drew out analytical frameworks that illuminate not just Intel but any dominant technology platform encountering a changed competitive context.

Burgelman’s engagement with Intel began in August 1988, when Grove came to Stanford GSB looking not to donate money but to teach. The deans didn’t know what to do with him and sent him to Burgelman. Grove arrived, as always, extraordinarily well prepared, with three proposed topics for Burgelman’s class: ASICs (“we’re no longer really doing that”), the DRAM exit (“we exited from the dynamic random access memory business”), and RISC (“but we have decided we’re not doing that”). Within months, Intel’s annual report featured three chips on its cover — one of them the i860, a RISC processor. When Burgelman pointed this out, Grove explained that an engineer named Les Kohn had told Gordon Moore and Grove he was building a co-processor for the 486. He had actually built a standalone processor and found customers in the workstation market. Up to 50% of Intel’s development effort had quietly migrated into RISC while the CEO thought they weren’t doing RISC.

Burgelman visited Intel to meet Kohn. In the lobby, he saw a young guy with gold pens. “That’s Les Kohn?” The guy who had effectively told Andy Grove and Gordon Moore something different from what he was actually doing — and they hadn’t noticed. Burgelman’s observation: “That’s the driving force in innovation. It’s usually not the CEO, it’s the people who are deeper, deeper in your organization.”

The Three Strategic Skills in Semiconductors

Burgelman opened with a taxonomy he considers foundational. Three capabilities determine competitive position in semiconductors:

Design is an intellectual effort — it lives on paper and in software, it is portable and capital-light.

Process technology is applied material science: how you take a design and put it into silicon. It determines line width, which drives both yield (smaller features are less likely to be hit by a substrate impurity) and performance (closer lines mean electrons travel shorter distances, faster). Intel was historically strong here.

Large-scale manufacturing is the capacity to achieve high yields early in production at volume. Intel was not particularly strong here — a fact that would matter when Japanese firms entered the DRAM market in the early 1980s.

Fujitsu, Hitachi, and other Japanese firms identified DRAM as strategically critical for consumer electronics and computing. They entered with manufacturing discipline and miniaturization capability that Intel couldn’t match. “When I was young in the 1980s,” Burgelman said, “it looked like the Japanese were going to take over the world.”

The DRAM Exit: How a Financial Rule Made a Strategic Decision

The story of how Intel exited DRAM is not a story of a deliberate decision at the top. It is a story of how a financial rule embedded in resource allocation eliminated a business line while senior leadership was still debating whether to exit.

The revenue picture Burgelman showed — drawn directly from Strategy is Destiny — covers 1972 to 1988. In 1972, memory products accounted for roughly 90% of Intel’s revenues. Logic (microprocessors) were a small and growing sliver. Through the late 1970s, memory stayed dominant. Then, beginning around 1982, the lines crossed and crossed back repeatedly — a jagged period Burgelman labeled the “reorientation” on his slide — three years during which Intel was simultaneously a memory company and a microprocessor company, depending on who in the organization you asked. By 1985–1986, logic crossed above 50% of revenues and kept climbing. Memory fell toward zero.

The mechanism behind that curve was Intel’s resource allocation rule: maximize margin per wafer start. On constrained fab capacity, the product with the highest margin gets priority. This sounds rational. The implication runs deeper.

Burgelman put the question to the class: if I apply this rule to my resource allocation, what does it tell me about how I think about what kind of company I am? Margin is not profit — profit is margin times volume. Maximizing margin says nothing about volume. It says something about identity. Louis Vuitton maximizes margin. Intel, by this rule, thought of itself as the Louis Vuitton of semiconductors — a leading-edge company, not a volume company. High margins were not a negotiated outcome; they were a right.

Now apply this rule in a commodity market. In a commodity, margins collapse by definition. The margin-per-wafer-start rule therefore diverted fab capacity away from DRAM — round by round — to microprocessors, which were a specialty product with high margins. The DRAM team, still believing themselves to be a leading-edge operation, kept trying to develop next-generation products to “bend the curve back up.” Each cycle, the same rule redirected capacity away from them. Intel’s DRAM market share fell from a majority position to roughly 3%.

Burgelman drew this as a graph for the class: fab capacity for DRAM on one axis, time on the other, the curve bending relentlessly downward despite the DRAM team’s periodic attempts to reverse it. “And down, and down, and down, and down until I have 3% market share. Why? Because of this. It’s extremely powerful.”

The R&D story compounds the problem. Gordon Moore’s justification for continued DRAM investment was the learning curve: unit cost falls with cumulative volume, so the highest-volume product drives the most process learning, and DRAM still had the highest unit volume. “Therefore,” Moore argued, “this is our technology driver.” The argument was right in principle. It was wrong in practice. Intel had 3% market share. Hitachi had 15%. “So who has the steepest learning curve?” Burgelman asked. The class answered: Hitachi. “So you’re fooling yourself. This is Gordon Moore.” The DRAM team was funding process R&D to stay competitive in a market they no longer controlled, while the competitor with the actual learning advantage was accelerating past them.

The actual exit decision is revealing in its own right. Burgelman and collaborator George Cogan found through interviews that Intel had decided in November 1984 never to bring the one-megabit DRAM into production — an effective exit made below the level of top management. When Burgelman updated his case study to reflect this date and gave it to Grove to teach with a week’s lead time, Grove pushed back: “November 84, why did you change this?” Burgelman held firm. Grove left. Two days later, he called back.

“Robert, I’ve done my own research and you guys are right. I actually got a very difficult decision.”

The CEO of Intel had not been fully aware of the moment his company exited its founding business. The formal announcement came in October 1985, when Grove flew to Intel’s Oregon facility and told the engineers: “Welcome to the mainstream of Intel — from now on, we’re a microprocessor company.”

The Microprocessor Platform: How Intel Won — and the National Stakes

The microprocessor did not come from a strategy. It came from a Japanese calculator company called Busicom that wanted to move from electromechanical to electronic calculators. They came to Intel for four or five chips. Ted Hoff, knowing Grove — then Intel’s president — would never approve a large resource commitment, reduced the scope to four chips. One was the 4004, the first microprocessor. Intel had sold the rights to it to Busicom for $60,000. “So that tells you they didn’t really think at the time this was going to be such a big deal.”

Hoff eventually convinced Intel’s leadership to offer Busicom price concessions in exchange for getting the rights back. At one point, a Japanese company held the rights to the Intel microprocessor. That fact surprised the class. It surprised Burgelman too when he learned it.

What transformed the microprocessor from a curiosity into a platform was the IBM PC. IBM chose Intel largely because of Intel’s technical marketing excellence — a clear product roadmap and a ferocious competitive campaign internally called “Crush the Competition,” targeting 2,000 design wins. The PC was win number 50 on their priority list. IBM created an installed base. Once nine of ten people on a team had PCs, the eleventh was told to get a PC. The installed base attracted independent software vendors, who increased the platform’s value, which reinforced the installed base. This is the mechanism Burgelman calls increasing returns to adoption.

The critical move came at the transition from the 286 to the 386. IBM had required Intel to cross-license the 286 to AMD, NEC (Japan), and Siemens (Europe). When Intel asked its licensees to share 386 development costs, they refused — counting on IBM to protect their position. Intel went to IBM and said it would not cross-license the 386. IBM resisted. But they had never demanded exclusivity as a condition of the cross-licensing agreement, which left a gap. Intel got Compaq to ship a 386-based PC. IBM’s marketing team, watching Compaq gain ground, told their engineers: “The train is leaving the station.” IBM had to follow.

Intel became sole-source supplier of the most important component in computing. As Burgelman’s slides note, that decision had consequences beyond Intel’s balance sheet: Intel’s sole-source strategy as of the 386 generation materially helped reverse the decline in the US’s global market share in microprocessors. The data from Strategy is Destiny (p. 136) tells the story precisely. In 1980, US companies held roughly 90% of worldwide microprocessor unit market share; Japan held roughly 10%. By 1985, the lines had nearly crossed — US share had fallen to around 50%, Japan had risen to match it. After Intel’s sole-source move locked in the 386 as the PC standard, US share recovered to roughly 60% by 1990 while Japan’s fell back toward 30%. What looked like a business decision was a national-security-relevant intervention in the global technology landscape.

“Intel Inside” was the mechanism for securing the capital investment this required — one of the few times in industrial history that a component manufacturer has successfully built a direct consumer identity, jumping over the OEM to the end buyer. Grove’s summary: “The PC is it!”

The computer industry had flipped from vertical to horizontal. IBM and DEC had once been vertically integrated. Now it was horizontal layers: chips (Intel, Motorola), operating systems (Microsoft), applications. The platform owners were Intel and Microsoft, not IBM.

Dominance, Co-evolutionary Lock-in, and the Prison of Moore’s Law

Intel’s sole-source dominance during the 1990s and into the 2000s made it what Burgelman’s framework calls a market-driving company — shaping the industry rather than responding to it. The clearest expression of this is Grove’s phrase: “The PC is it!” — a declaration that the PC was not just Intel’s biggest market but the market, the one that mattered, the one Intel would organize itself to serve and extend.

The problem with being market-driving is what happens next. Craig Barrett, who succeeded Grove as CEO, described the PC business as a “creosote bush” — a desert plant whose roots secrete a toxin that prevents anything else from growing nearby. The metaphor is precise. The PC business was not just Intel’s focus; it was actively suppressing the growth of everything adjacent to it. Burgelman’s formal term for the resulting condition is co-evolutionary lock-in: Intel and the PC industry had evolved together so tightly that Intel could no longer act independently of it.

His 2002 Administrative Science Quarterly article formalized this as a model. The diagram from that paper shows two channels through which strategy actually gets made in large organizations:

Induced strategic action follows the existing strategy. It flows through the organization’s tight structural context — the resource allocation rules, capital budgeting processes, incentive systems — and reinforces the current business strategy. In Intel’s case: allocate fab capacity to maximize margin per wafer start, fund Moore’s Law process advances, sell CPUs to PC OEMs.

Autonomous strategic action is bottom-up innovation that doesn’t fit the current strategy. Les Kohn building a RISC processor while claiming to build a co-processor is autonomous strategic action. The i860 project, WiMAX experiments, early AI chip efforts — all autonomous. For autonomous action to become strategy, it has to get legitimized through a strategic context — the process by which top management recognizes and institutionalizes an emerging opportunity.

In Burgelman’s diagram, the strategic context box is crossed out. That is the diagnosis. At Intel during co-evolutionary lock-in, autonomous strategic actions consistently failed to get legitimized. The tight structural context — the margin rule, the Moore’s Law discipline, the CPU-centric organizational identity — acted as two forms of inertia. Strategic inertia I prevented autonomous actions from making it through the structural context at all. Strategic inertia II prevented the strategic context from actually influencing the business strategy even when it existed. The result: emerging product-market environments (mobile, AI, communications) kept generating signals from within Intel, but those signals never became strategy.

Platform dominance creates the conditions for lock-in by design. The investments required to maintain market-driving status — Burgelman put fab costs at roughly $1–2 billion per cycle in the era he was describing, vastly more today — produce a disposition that is operationally useful right up until it becomes catastrophic. “You’re blessed,” he said, describing the feeling of being the world’s most essential component manufacturer. “And that, in a way, does it create arrogance? I think it eventually creates overconfidence, excessive self-confidence. And you no longer really pay attention to other things, because you’re so preoccupied.”

He was careful to resist the easy explanation. “It’s not just arrogant because I’m a bad person or something like that — that’s too easy an explanation. You have to try to understand what would be the drivers of that so-called error.” The Greek term for it is hubris. His illustration was Napoleon.

“When Napoleon went to Russia in 1812 — I read a lot of history and science and philosophy, not many business books. He wanted to fight the Russians before entering Russia. But because of Austerlitz in 1805, the Russians knew that would lead to defeat, so they withdrew. Napoleon has to decide: do I follow them or not?”

He traced this from a small memoir by an administrative general who recorded the meetings. The night before the invasion, Napoleon’s best marshals — Murat, Ney, the strongest and bravest — all said they couldn’t do it. They weren’t equipped. General winter would defeat them. Napoleon’s response:

“It’s not the rules that determine success. It is success that determines the rules.”

He went with 610,000 men. A student offered the inversion afterward: “You could say it’s failure that sets the rules.”

Moore’s Law as prison: because line width reduction continuously allows the CPU to absorb more capability, every new class of compute — graphics, communications, neural acceleration — can be answered with “we’ll incorporate it into the CPU eventually.” The incentive to build a dedicated architecture disappears. “They are trapped in the logic that has made them great. That’s what’s interesting. Arrogant — well, I’m going to find somebody else. How do you get out of that vector? The vector is great, as long as the context within which you are pursuing it is still the same. But what happens when that context changes? In that case, you are in the prison of Moore’s Law.”

Two Specific Failures: Communications and Mobile

Communications. Tom Lee raised Intel’s attempt to enter the communications business. Burgelman took it as confirmation of the hubris thesis. Grove had articulated a principle in the Wall Street Journal: “Volume begets standards. Not the other way around.” True in the PC industry. Not true in telecommunications, where standards are set by government-influenced committees.

Qualcomm would send 20 of their best engineers to standards committees, treating each meeting as a strategic investment. Intel would send two — “probably to save on travel expenses.” Intel had, as Burgelman noted with some bemusement, 5,000 PhDs in solid-state physics. The problem was not capability. It was strategic framing. Qualcomm understood committee presence as the mechanism for owning the platform. Intel treated it as overhead.

The consequences were concrete. In 1994, Intel launched ProShare, the first PC-based video conferencing product. It didn’t work — broadband was required, and at the time the only available option was ISDN. “I live in Portola Valley. I almost spent several thousand dollars on an ISDN line. I ended up not doing it.” Later came WiMAX, Intel’s attempt at wireless broadband — same structural problem, different decade. By the mid-2010s, as Burgelman’s slides recorded, “the mobile industry solidified around the low power UK-based ARM architecture, and the operators do not support the WiMAX standard.” Verizon said no. Burgelman wrote a case study on WiMAX too. “It’s the same thing.”

His summary: “Strategy is always about dependence versus influence. In the PC space, they had all the influence. In the telecommunications space, they had neither.”

Mobile. Apple was actively seeking Intel processors for mobile devices around the time of Paul Otellini’s tenure. The unit economics were roughly $30 per chip. Server-grade processors went for ten times that — “multiply by a factor of 10 if not more, right?” Intel, committed to its Louis Vuitton identity, said no to mobile. Keshavarzi, who worked at Intel designing chips during that era, described it as a deliberate strategic decision: “They chose to say no to it. How stupid is that? But honestly, honestly, it is related to what I just said.”

The deeper failure was Nvidia. Around 2005, Nvidia came to Intel seeking an acquisition. The price was $20 billion — “10 times price to sales ratio, very very very reasonable” — five years before the AI boom made that price a matter of historical comedy. Intel passed. Intel did eventually acquire Altera (FPGAs) and Habana and Nirvana (AI chips). Nothing came of any of them. “Intel’s got this habit of buying companies and not being able to productize anything out of it.”

The 2017 Virtuous Circle — and Its Unasked Questions

The restricted Stanford Business School case study Burgelman wrote on Intel in 2017 — the one Intel’s legal department has blocked from distribution outside his own class — centered on a diagram he called “Intel’s Virtuous Circle.” At the time, Intel believed it sat at the center of a self-reinforcing data flywheel: IoT devices (50 billion projected by 2020) generated data, that data flowed to the cloud and corporate data centers, where Intel’s architecture processed it, which drove more demand for Intel’s chips, which funded the Moore’s Law process advances that kept Intel ahead.

On the slide — which Burgelman showed in the 2020 symposium version — each node in that circle is annotated with a question rather than a claim. “How many on I/A?” (Intel Architecture). “How important is Intel?” “Does it matter? How to do it?” “How many from Intel? Which ones?” Every arrow in the supposed virtuous circle is followed by a doubt. The diagram is less a strategic plan than a set of unresolved bets. When Burgelman asked Krzanich what Intel was doing with AI in that period, Krzanich said the CPU plus FPGA could handle less well-defined algorithms where Nvidia’s GPU handled well-defined ones. Burgelman then went to Nvidia and asked Jensen Huang the same question directly:

Jensen said yes. But then he said: “Yes, but there are not enough software developers.”

Intel had 10,000 software developers. “But are they a software company? No. Just like Nokia was — they are a hardware company. It’s extremely difficult for a hardware company.”

Burgelman’s 2020 Diagnostic: The Questions Before Gelsinger

The slides Burgelman used in class were from a Stanford Engineering School symposium in October 2020 — the same symposium he mentioned having been asked to speak at. The final substantive slide posed a set of strategic questions about Intel’s future. Every one of them reads, in retrospect, as a precise forecast of what the Gelsinger era would attempt and fail to resolve.

Short-to-medium term:

1. Can Intel regain leadership in driving Moore’s Law? If so, under what conditions and when? If not, which company will take over the market-driving role?
2. Can the general purpose Intel Architecture continue to sustain the horizontal structure of the microprocessor industry in the face of “verticalization” efforts by: cloud services providers (the big 7); Apple replacing IA chips with its own silicon; Microsoft and Qualcomm’s ARM-based PC architecture; and Nvidia’s prospective acquisition of ARM?
3. How comfortable is the US government with strategically critical national security players dependent on TSMC?
4. If Intel becomes a market-driven rather than market-driving company, does it make sense to remain an Integrated Device Manufacturer?
5. If not, how could a freestanding Intel Foundry become able to compete with TSMC and Samsung, and in what time horizon?
6. How to support TSMC’s move to the US and develop a second source?

Gelsinger’s 2021 answer to question 4 was “yes, stay integrated.” His answer to question 1 was “we’ll catch TSMC by 2025.” His answer to question 5 was to create Intel Foundry Services. None resolved cleanly. The questions remain open — now under Lip-Bu Tan.

The Pat Gelsinger Era (2021–2024): Four Simultaneous Battles

By the mid-2010s, the industry dynamics Burgelman had been tracking had solidified against Intel. His slides from the 2020 symposium catalogued the state of play: mobile had locked in around ARM; Nvidia’s GPU was dominating AI and machine learning in the data center; in the foundry business, “Taiwan-based TSMC takes the lead in driving Moore’s Law and serves all the US-based fabless companies (including AMD). Intel disputes the linewidth nomenclature change used by TSMC but is for the first time no longer the leader.”

That last phrase — “for the first time no longer the leader” — marks the threshold. Gelsinger entered in early 2021 with four simultaneous battles to fight:

Manufacturing vs. TSMC. Intel had been stuck at the 10nm node for roughly five years before Gelsinger’s arrival. TSMC had moved ahead. Burgelman was in San Francisco listening to Intel’s senior leadership explain that TSMC’s claims of being ahead were a matter of “terminology” — that Intel had “changed the terminology” and wasn’t really behind. “So they were fooling themselves.” Intel’s capex as a percentage of revenue was approaching 50% — roughly equivalent to TSMC — but TSMC was deploying that capital while serving the entire global market. Intel was deploying it primarily to serve its own products.

X86 vs. AMD. AMD, using TSMC for manufacturing and sharing the x86 instruction set architecture under long-standing licensing terms, had closed and in some benchmarks exceeded Intel’s performance. AMD’s fabless model meant minimal capital expenditure relative to revenue. Intel was investing at TSMC-scale capex but capturing only a fraction of TSMC’s volume.

ARM. The ARM instruction set — low-power, royalty-based, widely licensed — had dominated mobile and was moving into laptops and edge infrastructure. Intel had declined the mobile market a decade earlier. ARM licensees were now occupying that space permanently.

Nvidia and AI. The GPU’s suitability for matrix algebra — the foundation of modern AI training and inference — had made Nvidia’s CUDA software ecosystem the default platform. Platform lock was as much about software as silicon.

The strategic geometry: “I have four opponents that all are benefiting from what I used to benefit. How the hell am I going to win? I’ve told you when I started, this is a sad story. I really think it is.”

Gelsinger’s response was to contest all four simultaneously: keep manufacturing, compete with TSMC on foundry services, compete with AMD on x86, compete with Nvidia on AI, build Intel Foundry Services. Each competitor already in a reinforcing flywheel, Intel’s capex consuming the balance sheet, and the software talent dispersed to Amazon, Apple, and Microsoft.

“In a world of winners take all, Andy Grove and I used to say there are only three strategies: be the biggest, become a niche player, or get out.”

The board and the activist investors. Malchow put the firing into legal context. Activist investors wrote what lawyers call a “fight letter.” The comparison doing the work was AMD: AMD had sold its fabs to GlobalFoundries, gone fabless through TSMC, and was outperforming Intel on product and market cap simultaneously.

What Comes Next: Lip-Bu Tan and the Open Questions

Intel’s market cap roughly doubled in the two years after Gelsinger’s departure. Burgelman was asked what Lip-Bu Tan’s vision is. His answer: “I don’t know.” [Laughter.] The recovery is driven by cost-cutting — roughly 20,000 layoffs, elimination of Intel’s dividend, and investor belief that Tan is a more disciplined operator. Tan’s declared position: skip the 18A node Gelsinger had committed to and intercept TSMC at the next generation. “Nothing nothing substantially has changed,” Keshavarzi said.

Tom Lee’s conjecture was the session’s most direct assessment:

“My guess — and this is just a random guess — is that he’s going to dismantle Intel. I don’t see any business way out of that. They can’t continue to fight a forefront war. It’s just not feasible. And so if you exclude that dream, then it leaves you with a very, very poor set of options. I don’t think the foundry business is going to succeed even with the government desperately trying to hang on to a domestic source of foundry capacity. I just think that’s pushing on string.”

Malchow’s preferred niche: inference compute. AI inference — running already-trained models at the edge and in data centers — is where the volume problem will be enormous, and Nvidia’s margins (~68–70%) are high enough that a well-positioned competitor can build a business there. “The idea that the constant build-out of both low power and data center inference is all going to be sold by Nvidia at Nvidia’s margins — that strikes me as something that someone is going to succeed in disrupting,” Lee agreed. “Certainly a fat target.”

Two M&A threads circulating at the time of the session: Qualcomm had reportedly considered acquiring Intel’s design operations (not the fabs — they are committed to TSMC and wouldn’t want the manufacturing liability). And Intel was in reported discussions to acquire SambaNova, a DARPA-originated, Stanford-based AI compute startup — notable because Lip-Bu Tan is currently chairman of SambaNova’s board. Keshavarzi’s uncertainty: “I don’t know whether that push is going — let me get rid of all of my investments and recover — or we really are going to go after something.”

The talent dispersal point: the engineers who built Intel are running design teams at the hyperscalers. Ronnie Borkar and others like him are corporate VPs at Amazon, Apple, Microsoft. Joe Malchow’s structural observation: we are in “an era where self-consumption of the component has frequently borne great rewards for your product and your share price” — Apple designing its own silicon, Amazon designing Trainium and Graviton, all manufactured at TSMC. The Intel-trained engineers make that possible. The engineering talent didn’t disappear; it redistributed to serve a new industrial architecture.

The TSMC-in-America Problem

The session closed on fab geopolitics. A student proposed that Intel’s unused Ohio fab capacity be sold to TSMC. Another student dismissed it: “TSMC knows that it is effectively a shield for Taiwan. There is no incentive for TSMC to purchase fab capacity in the United States because the moment semiconductor manufacturing leaves Taiwan and moves to the US, that’s a direct sacrifice of Taiwan’s national security.”

Burgelman added that a colleague had raised the inverse idea — why doesn’t the US government simply buy TSMC? He answered his own question: Taiwan’s national security calculus would prevent it.

On the Arizona fabs: Lee was characteristically concrete about the construction quality problem. In Taiwan, plumbers, fiber crews, and construction teams all coordinate; a single trench gets dug once and everything goes in together. “Here it’s like there’s no coordination at all. And if one group shows up and sees that some other group is gone, they go, well, in that case I’ll just go home.” Burgelman added the operational dimension: at TSMC, fab operators are on call 24/7. No holidays. “If there’s a problem with the fab, you get the call and you have to show up. They’re pumping wafers through it at fierce velocity. You cannot interrupt that.” Malchow noted a counterexample: Enphase moved a power electronics line from Guangzhou to Guadalajara and it outperformed its Chinese predecessor — cheaper too, mostly because trucks beat boats. The cultural gap is not insurmountable. It may be too large to bridge fast enough to matter for Intel’s current situation.

Burgelman’s 2020 slides were already asking the right question about this: how to support TSMC’s move to the US and develop a second source. His long-term policy recommendations from the same deck: (1) change immigration policies to attract the most talented people globally as PhD students in semiconductor-related fields — noting that Grove himself, and many other semiconductor leaders, were immigrants; (2) government support for academic research in semiconductor-related areas to create excitement and retain talent; (3) government support for applied research focused on speeding the translation of fundamental findings into practice.

Key Frameworks

Resource allocation rules encode organizational identity. Intel’s “maximize margin per wafer start” was not a financial formula — it was a statement about what kind of company Intel was. Any rule governing how capital and capacity are allocated will shape organizational identity and eventually limit strategic options. The DRAM exit was not a decision. It was the output of a rule applied consistently over time.

Middle management makes strategy through technical decisions. Grove’s own phrase is exact: “There were some middle-level managers who had made technical decisions already that limited the decision space of top management.” Les Kohn building a standalone processor while telling leadership it was a co-processor. Middle managers deciding in November 1984 not to bring the one-megabit DRAM into production. The locus of strategy is often not the executive suite.

Co-evolutionary lock-in: induced and autonomous strategic action. Burgelman’s formal model distinguishes two channels of strategy. Induced strategic action follows the current strategy through the tight structural context — resource allocation rules, budgeting systems, incentive structures. Autonomous strategic action is bottom-up innovation that doesn’t fit the current strategy. For autonomous action to become strategy, it must pass through a strategic context — senior management’s recognition and legitimization of an emerging opportunity. In Intel’s case, the strategic context was effectively blocked. Two forms of strategic inertia operated simultaneously: one preventing autonomous actions from clearing the structural context, another preventing the strategic context from influencing the business strategy even when it existed. The formal consequence: emerging product-market environments keep generating internal signals, but those signals never become strategy. Intel engineers saw mobile, saw AI, saw communications — and the organization could not act on any of it.

Increasing returns to adoption create self-reinforcing flywheels. Installed base → more developers → more platform value → larger installed base. This cycle is hard to break from outside. It is also, from inside, a mechanism for becoming “excessively self-preoccupied” — too focused on defending the flywheel to notice when the context around it changes.

The vector is a prison when context changes. Grove’s concept of vectorizing — aligning all organizational forces toward a clear strategic direction — is a source of competitive power and operational cohesion. It is also the reason Intel couldn’t see mobile, couldn’t see communications standards, and couldn’t see AI. “The vector is great, as long as the context within which you are pursuing it is still the same.”

On strategic leadership and shi**.** Burgelman’s final slide introduced a non-Western perspective on the context Intel now operates in. Henry Kissinger, writing on China, observed: “The traditional Chinese view of history emphasized a cyclical process of decay and rectification, in which nature and the world can be understood but not completely mastered. The best that can be accomplished is to grow into harmony with it. Strategy and statecraft become means of ‘combative coexistence’ with opponents. The goal is to maneuver them into weakness while building up one’s own shi, or strategic position.” The concept of shi derives from Sun Tzu’s Art of War. Burgelman’s implication: strategic leaders facing a multi-front competitive environment need to understand the strategic logics their opponents are operating from — not just optimize their own vector.

Burgelman’s primary works on Intel: Strategy is Destiny (2002, foreword by Andy Grove); the 2002 Administrative Science Quarterly article “Strategy as Vector and the Inertia of Co-evolutionary Lock-in.” His restricted 2017–2018 Intel AI case study is Stanford Business Case SM-286. His current seminar at Stanford GSB is S573. His co-teacher is George Cogan, former Bain partner, who appears in several of the Intel anecdotes in this session. The slides used in class were originally presented at the Stanford Engineering School symposium “Semiconductor Technology and Manufacturing in the US in the 21st Century: Back to the Future for Lessons from the Past,” October 16, 2020.