Chapter 9
What makes a strategy and a structure fit — and what happens when they don't?
In December 2024, the board of Intel — for half a century the definitive American technology company — forced out its CEO. Pat Gelsinger's four-year turnaround attempt had ended in a $16.6 billion annual loss, a suspended dividend, a stock down 60 percent, and the announcement of 15,000 layoffs; the board, sources said, had lost confidence in his plan to answer NVIDIA's dominance.12 Intel had spent those years pursuing everything at once: designing chips, manufacturing its own chips, manufacturing other companies' chips as a new foundry business, and building AI accelerators — an empire-scale strategy run through an empire-scale structure of roughly 125,000 people and many management layers.34 His successor, Lip-Bu Tan, arrived in March 2025 with a diagnosis that was equal parts strategy and structure: Intel had "invested too much, too soon, without adequate demand," its factory footprint had become "needlessly fragmented," and the organization needed fewer layers, fewer people — headcount would fall below 100,000 — and "no more blank checks."45
About ten miles away sits the company that ate Intel's future. NVIDIA — by 2025 the most valuable company in the world, worth over $4 trillion — is run through an organizational design most textbooks would flag as malpractice: CEO Jensen Huang has roughly 60 direct reports and holds one-on-one meetings with none of them. "I don't have one-on-ones with them because it's impossible," he says. "We present a problem, and all of us attack it." He calls the method "extreme co-design," gives feedback publicly so everyone learns from it, and openly mocks the standard "hamburger" org chart — leadership on top, middle management in the middle — as making no sense.678 The structure serves a strategy of unusual focus: one accelerated-computing architecture, compounded for nearly two decades. Huang recalls that carrying the CUDA software platform on every consumer chip once consumed all the company's gross profit and helped sink its market value to about $1.5 billion — a bet the layered giants of the industry declined to make.9
The tempting headline — flat beats tall, focus beats empire — collapses on inspection. Flat structures fail constantly; sixty direct reports would destroy most CEOs and most strategies; and Intel's wounds trace at least as much to strategic choices made decades ago (passing on mobile, passing on AI) as to its layer count. Meanwhile Intel's 2025 recovery — six straight earnings beats, an 84 percent stock rebound, a U.S. government stake, and a $5 billion investment from NVIDIA itself — came from shrinking both the strategy and the structure at once, which rather scrambles the question of which one was the disease.510 So which is cart and which is horse? Does structure follow strategy, strategy follow structure — or do they fail and succeed only as a matched pair?
The central question: what makes a strategy and a structure fit each other — and what happens when they don't? The tools: what strategy actually is (positions and resources); Chandler's famous thesis that structure follows strategy, and the basic structural forms; the mechanistic–organic contingency that ties design to environment; the design dials every organization sets — span, centralization, formalization; ambidexterity, the art of exploring and exploiting at once; and what platform strategies and AI are doing to structure right now. Intel and NVIDIA then return as the decade's defining matched pair.
Strategy is an organization's theory of how it will win: which customers, which activities, and — as important — which not. The field's two great lenses answer "where does advantage come from?" differently. Michael Porter's positioning school locates it outside the firm, in industry structure and generic positions — cost leadership, differentiation, or focus — and warns that firms "stuck in the middle," pursuing all at once, get outcompeted at each.12 Jay Barney's resource-based view locates it inside: sustained advantage flows from resources that are valuable, rare, inimitable, and organized for exploitation (VRIO) — brands, capabilities, ecosystems that rivals cannot buy or copy.13 The lenses converge on this chapter's purposes at one point: both make strategy a discipline of choice and sacrifice, and both imply structural consequences — a cost leader needs standardization and tight control; a differentiator needs creative slack; a rare capability (say, a software ecosystem two decades deep) needs an organization built to compound it rather than harvest it. Limits: positioning can understate how firms change industries rather than pick spots in them, and VRIO reasoning turns circular when "valuable resources" are identified only after the winning.
Organizational structure is the formal system dividing work and coordinating it back together — visible in the org chart, enacted in reporting lines, spans, and decision rights. Alfred Chandler's landmark historical study of DuPont, General Motors, Sears, and Standard Oil produced the field's most famous sentence: structure follows strategy. Firms that diversified while keeping centralized functional structures drowned in coordination overload until they invented the multidivisional form — and "the most complex structures result from the concatenation of several basic strategies."11
• Functional — organized by specialty (engineering, sales, finance): deep expertise and scale economies; slow cross-functional coordination; fits focused, single-business strategies.
• Divisional — organized by product, market, or geography: accountability and speed per business; duplication and divisional silos; fits diversified strategies.
• Matrix — dual reporting across function and product/project: flexibility and resource sharing bought with role conflict and slower decisions.
• Network / platform — a small core orchestrating partners and ecosystems: reach without headcount, at the price of control.
Chandler's thesis remains the starting diagnostic: when strategy changes and structure doesn't, the mismatch shows up as overload, delay, and internal warfare — Intel simultaneously running a products business and a customer-serving foundry inside one integrated structure is a modern sitting for the portrait, and the eventual move to separate the foundry as a subsidiary is Chandlerian correction in real time.13 The honest amendment the field added later: causation also runs backward — structures shape what strategies an organization can perceive and execute (a firm of layers literally cannot see what a firm of sixty co-designing experts sees weekly), so the pair co-evolve. Treat "structure follows strategy" as a design obligation, not a law of physics.
The contingency tradition asked what environments demand of design. Burns and Stalker, studying firms entering the fast-moving postwar electronics industry, found two coherent systems: mechanistic structures — tall hierarchies, precise roles, vertical communication, knowledge concentrated at the top — that excel in stable environments, and organic structures — fluid roles, lateral communication, expertise-based authority — that fit turbulent ones.14 Lawrence and Lorsch added the refinement managers actually use: successful firms in complex environments are differentiated (subunits as different as their sub-environments require — research organic, manufacturing mechanistic) and, critically, invest matching effort in integration — the liaison roles, teams, and systems that stitch the differentiated parts back together; high performers differentiated and integrated more than rivals.15 Why it matters: this is the theory that forbids copying NVIDIA. An organic, sixty-report, public-feedback design fits a single-architecture strategy in a hyper-turbulent industry staffed with elite experts; imposed on a utility, a hospital system, or a five-business conglomerate, the same design produces chaos, not speed. Limits: environments are partly chosen and partly constructed by firms (NVIDIA made its environment turbulent for everyone else), and the mechanistic/organic poles are ideal types — real organizations are portfolios of both, which is Section 2.5's subject.
Every structure is a setting of three dials, each a trade-off rather than a virtue. Span of control (Week 1's variable, now in context): narrow spans buy supervision and coaching at the cost of layers; wide spans delete layers — and every deleted layer removes a filter that distorted information on its way up and slowed decisions on their way down, which is Huang's explicit argument for sixty reports and no hamburger.68 Wide spans only work when subordinates are expert and self-directing, and when the leader substitutes broadcast mechanisms (group problem-solving, public feedback, shared context) for one-on-one supervision — NVIDIA's "extreme co-design" is precisely that substitution, not the absence of management.69 Centralization: concentrated decision rights buy consistency and strategic coherence at the price of local responsiveness and frontline motivation (Week 3's autonomy); decentralization reverses the trade. Formalization: rules and standard procedures buy reliability, fairness, and scale at the price of adaptation — the mechanistic dial. The craft is coherence: the dials must agree with each other and with the strategy. Intel's crisis pattern — high formalization, many layers, and a strategy demanding rapid pivots into AI and foundry services — is what dial-incoherence looks like at scale; Tan's simultaneous moves (fewer layers, fewer people, narrower strategy, "no more blank checks") reset all the dials toward a coherent, smaller machine.45
James March named the deepest tension in organizational design: exploitation (refining what you already do well — efficient, reliable, self-reinforcing) versus exploration (searching for what's next — inefficient, failure-prone, and the only source of renewal). His central warning: because exploitation's returns are faster and surer, adaptive processes systematically crowd out exploration, and organizations become better and better at things that matter less and less.16 O'Reilly and Tushman's structural ambidexterity is the standard remedy: house exploratory units separately — their own processes, metrics, and cultures — while integrating them at the senior-team level so the new can draw on the old's assets without being smothered by its logic.17 The concept explains both hook companies at once. Intel is March's warning incarnate: a magnificent exploiter of x86 economics whose processes filtered out mobile and then AI until the environment collected the bill.13 NVIDIA is the rarer success pattern: it funded CUDA — a decade-plus exploration that consumed the company's profits and crushed its market value to $1.5 billion — inside the exploitation engine (GeForce carried it to the world), protected by founder conviction rather than structural separation.9 Limits: ambidexterity is easy to preach and brutally hard to govern (the exploit business always has better numbers this quarter), and founder-protected exploration à la NVIDIA is not a reproducible design — it is key-person risk that happened to pay.
Two contemporary forces are re-setting the dials economy-wide. Platform and ecosystem strategy: NVIDIA's deepest moat is not a chip but CUDA's ecosystem — millions of developers, tools, and libraries whose switching costs compound (a VRIO resource par excellence), and whose cultivation requires structure that treats outside developers as part of the organization's design.913 Competing with a platform demands either a better platform or an ecosystem coalition — a fact that turns strategy increasingly into architecture. AI-driven flattening: as Week 1 established, generative AI absorbs precisely the coordination, monitoring, and reporting work that justified management layers, enabling wider spans and fewer levels — the Harvard Business School evidence on 50,000+ developers shows individual contributors taking on former managerial tasks from below.18 Intel's cuts, Bayer's delayering (Week 1), and NVIDIA's native flatness are one pattern seen from three directions. The caution the whole chapter has been building: flattening is a capability bet, not a cost trick — it works where expertise, shared context, and broadcast leadership substitute for supervision, and it fails, expensively, where they don't. Structure is strategy's delivery system; thinning the delivery system of a strategy that still needs it is how companies mistake amputation for fitness.
The pairing holds the industry, the region, and even the decade's technological shock constant, and varies the strategy–structure pair. If either strategy or structure alone decided outcomes, the cases would be less instructive: Intel's strategy was broad where NVIDIA's was focused, and Intel's structure was layered where NVIDIA's was radically flat. The analytical work is showing that each company's outcomes flow from the fit — and the misfit — between the two, exactly as the contingency tradition predicts.
The facts: entering the 2020s, Intel ran an integrated device manufacturer strategy — designing and fabricating its own chips — layered with Gelsinger's 2021 additions: a foundry business to manufacture for others, subsidized fab construction across several states, and an AI accelerator effort chasing NVIDIA. The structure carrying it was the classic mechanistic giant: ~125,000 people, deep hierarchy, high formalization.34 The strategy demanded organic speed (foundry customers, AI roadmaps) from a machine built for exploitation reliability; the bill arrived as a $16.6 billion 2024 loss including $15.9 billion in impairments, a halted dividend, 15,000 layoffs, and Gelsinger's December 2024 ouster after the board lost confidence in the turnaround.123 Tan's 2025 reset moved strategy and structure together: strategic narrowing ("invested too much, too soon"; Ohio fab slowed; foundry discipline), structural compression (headcount toward under 100,000, layers removed, "no more blank checks"), and external re-anchoring — a U.S. government stake near 10 percent, $2 billion from SoftBank, $5 billion from NVIDIA, and by January 2026 the 18A process reaching volume manufacturing with Microsoft and Amazon as foundry customers.4510 Six consecutive earnings beats and an 84 percent 2025 stock recovery followed — though the company still posted a small 2025 loss, and analysts warn the rally has "baked in" a turnaround not yet delivered.3510 As a theory test: the decline is Chandler plus March — structure lagging strategy while exploitation logic filtered out two successive platform shifts; the recovery-so-far is the matched-pair prescription executed under duress. Honest caveats: much of the 2025 rebound reflects government backing, partner investments, and cycle timing rather than proven design fitness, and the deepest strategic wounds (mobile, AI) were inflicted by boards and CEOs long before anyone in this case held the job.1
The facts: NVIDIA's strategy is focus compounded — accelerated computing on one architecture, with CUDA as the ecosystem moat, carried through a near-death experience (gross profits consumed, market value down to ~$1.5 billion) because the founder believed the platform would matter.9 The structure is the strategy's mirror: Huang's 60-plus direct reports (55 at his 2024 Stanford telling), no one-on-ones, no hamburger layers, problems attacked collectively in the open, feedback delivered publicly so sixty people learn from one correction, and near-universal engineering fluency at the top table — "extreme co-design" as an operating system.678 The design's function is informational: with no filtering layers, the CEO and the expert leads share one context at the "speed of light," which is precisely what a single-architecture strategy in a hyper-turbulent market requires — every part of the stack (silicon, networking, software) must move together, so the structure fuses them at the top rather than coordinating them through middle layers.68 The result sits atop the market: the world's most valuable company, past $4 trillion.7 Honest caveats, each one a design cost: the structure is saturated with key-person risk (it runs through one founder's bandwidth and judgment); it presumes elite, self-directing experts and a strategy narrow enough for one table to hold — neither of which most firms have; Huang himself describes decades of feeling "30 days from going out of business," a personal cost the org chart quietly relies on; and a design this fused to one bet has never yet been tested by the bet going wrong.79
| Intel | NVIDIA | |
| Strategy | Broad: IDM + foundry + AI catch-up, pursued simultaneously | Focused: one architecture, CUDA ecosystem compounded ~20 years |
| Structure | Mechanistic giant: ~125K people, deep layers, high formalization | Radically organic: 60 direct reports, no 1:1s, public feedback, no middle layers at top |
| Fit diagnosis | Organic-demanding strategy in a mechanistic machine — misfit | Turbulence + expert workforce + narrow strategy — fit |
| March's tension | Exploitation crowded out exploration; mobile and AI filtered away | Exploration (CUDA) protected inside exploitation (GeForce) by founder conviction |
| Recent trajectory | $16.6B 2024 loss; CEO ousted; 2025 reset of strategy AND structure; fragile recovery | Most valuable company (> $4T); design untested by failure of the core bet |
| Transferability | Cautionary pattern broadly applicable | Low: presumes elite experts, one bet, one extraordinary founder |
Three conclusions, with the discipline the cases deserve. First, the pair confirms fit, not form: Intel did not fail because tall structures are bad, and NVIDIA did not win because flat ones are good — each outcome tracks the match between the structure's information-processing capacity and the strategy's demands, which is the contingency tradition's whole point.1415 Second, the cases run Chandler in both directions: Intel's structure lagged its strategy (classic), but its structure also shaped its strategy for decades — the layered exploiter could not perceive the platform shifts, March's crowding-out operating through the org chart itself.1116 Third, beware the survivor's template: NVIDIA's design is visible because its bet paid; the identical structure attached to a failed architecture would today be a cautionary tale about key-person dependence and missing guardrails, and Intel's 2025 rebound — built partly on government stakes and rivals' investments — is a recovery in progress, not a proof.510 The transferable question is the one every case in this chapter answers differently: what must this strategy's structure be able to see, decide, and coordinate — and is it built to do that?
Return to the central question: what makes a strategy and a structure fit — and what happens when they don't? Four conclusions carry.
Strategy is sacrifice; structure is its delivery system. A strategy that chooses nothing needs a structure that can do everything — which no structure can. Diagnose misfit by its symptoms: chronic overload, decisions aging in queues, and internal wars over resources are structural bills for strategic indecision.1112
Set the dials to the environment, and to each other. Mechanistic for stability, organic for turbulence, differentiated-and-integrated for complexity — and never copy a design without importing its conditions: NVIDIA's sixty-report flatness is a system of expert workforce, broadcast leadership, and narrow strategy, not a spans policy you can adopt on Monday.61415
Govern the explore/exploit boundary on purpose. Exploitation will always have better numbers this quarter; renewal dies quietly unless exploration is protected — by structure, by senior-team integration, or (least reliably) by founder conviction. Ask of any organization: where does the next business live, and who defends its budget?1617
Treat flattening as a capability bet. AI is deleting the informational work that justified layers, and the flattening wave is real — but every removed layer must be replaced by expertise, shared context, and leaders who can manage by broadcast. Where those substitutes exist, layers were tax; where they don't, layers were load-bearing. Knowing which is the design judgment this chapter exists to train.4618
Notes appear as superscript numbers in the text and correspond to the numbered sources above. DOIs are provided where available; classic books are cited to their original publishers.