MBUS 873 — Session 7 · Capstone

How Companies Compete in a Geopolitically Fragmented World

Queen's Smith AMBA 2026 · Prof. David Detomasi · Participation-Ready Prep
Export Controls Corporate Strategy Under Fragmentation US-China Semiconductor Standoff Balancing Act: Nvidia (HBS 9-725-009) Course Capstone
Block 1 — The Lens: Corporate Strategy in Geopolitical Fragmentation

From Country and Industry to Firm: The Capstone Move

Sessions 1–6 analyzed competitiveness, political risk, trade politics, and great-power rivalry from the vantage point of countries and industries. Session 7 pulls the lens down one more level: given all of that — given export controls, industrial policy, and a bifurcating US-China technology order — how should a single firm actually act? Nvidia is the sharpest possible test case: the company that makes the GPUs powering the global AI boom, caught directly between its home government's national-security mandate and its dependence on the world's second-largest datacenter market. As CEO Jensen Huang put it bluntly when asked about losing China: "If we are deprived of the Chinese market, we don't have a contingency for that. There is no other China, there is only one China."

Firm Strategy = f(Compliance, Market Access, Competitive Position)
Every corporate response to geopolitical fragmentation is a trade-off across these three variables — no option optimizes all three at once.

When Do Export Controls Actually Work? Four Conditions

Drawing on the case's own account of the U.S. export-control debate — Willy Shih's "thermodynamics" framing (controls slow the rate of diffusion, they don't stop it), the Griffin/Porter view that leadership and standard-setting beat containment, and the Center for a New American Security's critique of using national-security tools for industrial policy — four conditions determine whether an export control regime succeeds rather than backfires:

1

Technology Specificity

Is the controlled technology a narrow, identifiable chokepoint (like EUV lithography) or a diffuse general-purpose capability (like "AI")? Narrow chokepoints are enforceable; broad categories invite endless redefinition and workaround chips.

2

Enforcement Capacity

Can the regulator actually verify end use and close loopholes — Entity Lists, Unverified Lists, license requirements on third countries — faster than the target and its suppliers can route around them?

3

Allied Coordination

Are the other choke-point holders (Japan's tool makers, the Netherlands' ASML, Taiwan's TSMC) moving in lockstep? A unilateral U.S. control just hands revenue and market share to un-restricted competitors.

4

Difficulty of Substitution

How fast can the target build or buy a domestic alternative? If substitution is slow, controls buy real strategic time. If it's fast, the control accelerates the very self-sufficiency it was meant to prevent.

The Corporate Response Menu

Facing a control regime it cannot change unilaterally, a firm like Nvidia has four broad categories of response — and the case shows Nvidia has tried all four, with mixed results:

RESPONSE 1

Localization

Build compliant, "de-contented" products for the restricted market — Nvidia's A800, H800, and H20 chip variants, engineered to sit just under whatever performance threshold Commerce sets.

RESPONSE 2

Dual Supply Chains

Separate the compliant product line from the flagship line, or shift the revenue mix toward services (DGX Cloud) instead of hardware — so a single ban doesn't sever the whole relationship.

RESPONSE 3

Regulatory Arbitrage

Route around restrictions via third countries, cloud-service reclassification, or timing sales ahead of new rules — the exact behavior Raimondo warned Nvidia against, and the exact gap the October 2023 rule's 40-country license expansion was designed to close.

RESPONSE 4

Political Engagement

Lobby directly, and coordinate through industry bodies like the Semiconductor Industry Association, to shape how aggressively and how unilaterally the rules get written and enforced.

The Nvidia dilemma in one sentence: comply with a home government that treats advanced AI chips as a national-security chokepoint, while serving a global market where nearly a quarter of datacenter revenue once came from one country, without ceding that country's AI infrastructure buildout to a fast-rising set of state-backed domestic rivals.
Block 2 — Case Analysis: Balancing Act: Nvidia's Strategy in the US-China Semiconductor Standoff (HBS 9-725-009)
Case Summary

Nvidia — founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem to bring 3D graphics to consumer PCs — became the essential infrastructure company of the AI era. Its 2006 CUDA software layer turned graphics processing units (GPUs) into general-purpose parallel computers; the 2012 AlexNet breakthrough and the 2022 release of OpenAI's ChatGPT then turned that parallel-computing capability into the engine of the deep-learning boom. The results were extraordinary: fiscal year 2024 revenue grew 126%, the company's market cap crossed $2.2 trillion (third most valuable company in the world), and its latest H200 chip was rumored to start at roughly $41,000 — with Nvidia claiming it now addressed a $100 trillion total market. That growth made Nvidia both a Wall Street darling and a Washington target, because the same GPUs that train chatbots can also, per a National Security Council argument about "dual-use" technology, accelerate hypersonic-missile modeling and nuclear-weapons simulation — a Wall Street Journal investigation found China's top nuclear weapons research institute was a repeat buyer of U.S. GPU chips.

Beginning in October 2022, the U.S. Bureau of Industry and Security (BIS) required a license to export advanced AI chips to China, blocking sales of Nvidia's flagship H100. Nvidia's response was to engineer de-contented, China-specific variants — the A800 and H800 — that met the letter of the restriction while preserving as much performance as possible. Commerce Secretary Gina Raimondo made clear this cat-and-mouse strategy had a short runway: "If you redesign a chip around a particular cut line that enables them [China] to do AI, I'm going to control it the very next day." She followed through: the October 17, 2023 revision to the Export Administration Regulations banned the A800 and H800 outright, added new performance-density thresholds, extended license requirements to 40 additional countries to close third-party transshipment routes, and effectively meant any further compliant chip Nvidia could design would perform only around the level of its 2017-generation V100 — commercially uncompetitive.

The stakes are what make this Nvidia's "balancing act." China represented 20–25% of Nvidia's datacenter revenue as recently as the year before the case's writing; that share fell to roughly 9% under the tightening restrictions. The October 2023 announcement alone wiped out $50 billion in Nvidia market capitalization in a single day and put an estimated $5 billion of pending China shipments in limbo. CFO Colette Kress warned that prohibiting AI-chip sales to China would cause a "permanent loss of opportunities for the U.S. [semiconductor] industry to compete and lead" — because every dollar and every AI workload Nvidia cedes in China flows to state-backed domestic rivals (Huawei, Biren, Moore Threads) that are closing the performance gap far faster than most Western analysts predicted. By early 2024, Nvidia was reportedly readying a further-compliant H20 chip for China but had not yet secured export licenses to sell it — leaving Huang to navigate export compliance, market access, and competitive position simultaneously, with no clean option that satisfies all three.

Why This Case Closes the Course

Nvidia sits at the exact intersection of every framework the course has built. It is a national-competitiveness story (Session 1) about whether the U.S. semiconductor ecosystem can out-innovate rather than merely out-restrict China. It is a political-risk story (Session 2) about regulatory risk originating not abroad but at home — Nvidia's biggest threat is its own government's Commerce Department. It is a trade-politics story (Session 3) about export controls as the sharpest tool in the protectionist toolkit. It is the direct continuation of the US-China rivalry and semiconductor cases (Sessions 4–5) — Taiwan's TSMC chokepoint is the manufacturing reality underneath every chip Nvidia designs. And it forces the question every prior session has circled but never had to fully answer: when a government and a global market pull a firm in opposite directions, what should management actually do?

Block 3 — When Are Export Controls a Good Idea? (Case Question 1)

Export Controls Work When They Slow Diffusion — Not When They Try to Stop It

The case's own history lesson is the sharpest evidence here: U.S. controls on cryptography during the Cold War were "traditionally used to restrict the spread of sensitive technology for national security objectives," per the Center for a New American Security, but by the 1990s the government recognized that "attempting to control the export of mathematics was futile" and pivoted from containment to leadership and standard-setting — as Michael Griffin and Lisa Porter (former Defense Department officials) put it, "it is more effective to control how technology is used by being the leader in development and promulgation of standards globally, rather than trying to contain it." That pivot let the U.S. shape the "standards and protocols used in [the] global implementation" of encryption while still capturing the economic boom that followed from e-commerce. This is the model export controls should aim for: narrow enough to be enforceable, targeted enough to buy real time, and paired with a leadership strategy rather than a pure containment strategy.

Applying that lesson, export controls succeed when most or all of the four conditions from Block 1 are present:

  • Narrow, identifiable chokepoint — not a broad category like "AI" that invites redefinition.
  • Real enforcement capacity — the ability to verify end use and close loopholes before they're exploited at scale.
  • Genuine allied coordination — every other holder of a chokepoint technology (fabrication, lithography, EDA tools) restricting in parallel, not years apart.
  • A target that cannot quickly substitute — where domestic development realistically takes years, not months.
Where controls backfire: The Diplomat's account of the 2022–23 restrictions is the cautionary case: "companies in Japan and the Netherlands initially continued to sell semiconductor manufacturing equipment that enabled the very capabilities Washington sought to restrict. U.S. businesses lost out on revenue and suffered reputationally, while Chinese firms still got their machines. When Japan and the Netherlands finally agreed to implement mirroring controls, China bulk-imported the machines right up until the restrictions went into effect." A control regime that isn't coordinated at the moment it's announced doesn't just fail to bite — it actively disadvantages the firms of the country that imposed it first.

Bottom line for Q1: export controls are a good idea only when the technology, the enforcement apparatus, and the allied coalition are all narrow and aligned enough to actually slow diffusion — and even then, they buy time rather than permanent advantage. Used as a blunt instrument against a broadly diffusing capability, or announced unilaterally ahead of allied partners, they risk handing revenue and market share to both the target country's domestic champions and the exporting country's own uncontrolled competitors.

Block 4 — Are Advanced Semiconductors a Good Candidate for Export Controls? (Case Question 2)

My Position: A Better Candidate Than Most — But the October 2023 Execution Outran the Conditions

Applying the four conditions from Block 1 directly to advanced AI semiconductors:

1. Technology Specificity — Strong Case

The Center for Strategic and International Studies identified four genuine chokeholds the U.S. controls target: chip exports, EDA design software (a near-monopoly held by three U.S. firms — Synopsys, Ansys, Cadence), semiconductor manufacturing equipment (ASML's EUV lithography monopoly), and U.S.-built semiconductor components. CSIS's Gregory Allen called this "an astutely targeted step forward for U.S. national security and technology leadership." Unlike a broad ban on "AI," this is about as narrow and identifiable as a chokepoint gets.

2. Enforcement Capacity — Mixed

BIS has real tools (Entity List, Unverified List, expanded license requirements to 40 countries in October 2023 to close transshipment routes), but a follow-up CSIS report less than two years after the initial 2022 controls found "a mixed picture" with "gaps in compliance between U.S. companies and those of allies" — an admission the enforcement apparatus was still catching up to the policy's ambitions.

3. Allied Coordination — Late, Not Absent

Japan and the Netherlands eventually adopted mirroring controls, but the lag let China stockpile — Chinese imports of chip manufacturing equipment rose 14% to over $40 billion in 2023, much of it pre-positioning ahead of tighter rules. Tokyo Electron, a Japanese toolmaker not subject to the same restrictions as Lam Research and Applied Materials, raised its revenue guidance 11% citing stronger-than-expected Chinese demand — direct evidence of controls redirecting revenue to non-restricted competitors rather than eliminating it.

4. China's Substitution Capacity — The Weakest Link

This is where the case most complicates the "controls will work" thesis. Just 11 months after the October 2022 restrictions, Huawei launched the Kirin 9000S, built on SMIC's 7nm process — matching Intel's then-current node — without access to ASML's EUV lithography or U.S. EDA software, using an older deep-ultraviolet, self-aligned quadruple-patterning workaround. Huawei's Ascend 910B chip, which Baidu ordered 1,600 units of in Q4 2023, tested on par with Nvidia's H100; Baidu's CEO claimed its own chatbot was "not inferior in any aspect to GPT-4." State-linked investors put $7 billion into memory-chip maker YMTC. None of this means China has closed the gap at the leading edge — but it is closing faster than the 2022 CSIS "comprehensive and permanent" framing assumed.

11 mo.
From Oct. 2022 restrictions to Huawei's Kirin 9000S (7nm) launch
$40B+
Chinese chip-equipment imports in 2023, up 14% — pre-restriction stockpiling
14%
DRAM/logic node threshold (16/14nm) set by Oct. 2022 BIS rule
40
Additional countries added to license requirements, Oct. 2023, to close loopholes
Counterargument I address directly: some analysts (Gregory Allen, CSIS 2022) argue the controls remain a durable win because "the gap in performance will grow over time as the world advances and China remains stuck behind." I disagree with the "permanent" framing specifically — Willy Shih's thermodynamics analogy is the more accurate model. Controls change the rate of diffusion, not whether diffusion happens. Advanced semiconductors pass conditions 1 (specificity) reasonably well, but the case's own evidence on conditions 3 and 4 — lagging allied coordination and faster-than-expected Chinese substitution — means the October 2023 rules are likely buying the U.S. a multi-year lead rather than permanent containment, at the real cost of ceding near-term Chinese AI-infrastructure revenue and ecosystem learning to Huawei and its peers.

Bottom line for Q2: yes — advanced semiconductors are a good candidate for export controls, better than most technologies, because the chokepoints (EUV, EDA, leading-edge fabrication) are genuinely narrow and hard to replicate quickly. But "good candidate" is not the same as "guaranteed success," and the October 2023 rules moved faster and more unilaterally than the enforcement and allied-coordination conditions could support — which is precisely the gap Nvidia has to navigate in Block 5.

Block 5 — How Should Nvidia Respond? (Case Question 3)

The Real Options on the Table

The case lays out — and history since has largely confirmed — five paths, each trading off compliance, market access, and competitive position differently:

OPTION 1

Comply Fully, Cede the Market

Accept a shrinking China revenue line and focus entirely on the U.S. and allied hyperscaler boom (already visible in Nvidia's Compute & Networking segment growth). Safest legally and politically — but Huang's own words ("there is no other China") signal no market replaces it at this scale, and it hands China's AI buildout entirely to domestic rivals.

OPTION 2

Keep Building Cut-Down Chips

Continue the A800 → H800 → H20 progression. The case shows this door keeps closing — Raimondo's explicit "the very next day" threat, and the October 2023 rule's next-generation compliant chip topping out near 2017's V100 performance. Diminishing commercial returns with each iteration.

OPTION 3

Shift to Compute-as-a-Service

Scale DGX Cloud (launched March 2023 with Oracle) instead of selling hardware outright — precedent exists in Microsoft Azure's 700% six-month demand growth in Hong Kong. Retains revenue and gives Nvidia visibility into end use, but Beijing's Data Security Law creates its own friction, and Washington was reportedly already weighing cloud-access restrictions too.

OPTION 4

Double Down on Political Engagement

Lobby directly and via the Semiconductor Industry Association. Track record is weak — SIA and chipmakers spent over $100 million ahead of the CHIPS Act but failed to stop the October 2023 BIS rules, and the industry is no longer unified now that Intel's CHIPS subsidies align it with the pro-restriction camp.

OPTION 5

Diversify the Revenue Base

Treat China as a bounded, managed exposure rather than a growth pillar, and let the non-China hyperscaler and enterprise AI buildout — already the larger and faster-growing part of the business — become the structural center of gravity.

My Recommendation: Comply Substantively, Lead With Services, Redirect Political Capital

Nvidia should stop treating Option 2 (redesigning around the letter of the rule) as a viable long-term strategy — Raimondo's warning was explicit, and the case shows the regulatory ratchet closing faster than Nvidia's engineering cycle can outrun it. Instead:

  • Comply substantively, not just technically — avoid the "perceived flouting" of regulations that drew the Commerce Department's rebuke over the H20. Every enforcement action taken against Nvidia specifically (rather than the industry broadly) increases political risk to the far larger non-China business.
  • Make DGX Cloud the primary vehicle for whatever legal China exposure remains — better margins, better end-use visibility, and structurally harder for a Chinese buyer to reverse-engineer or diffuse than a physical chip sale.
  • Redirect political engagement away from fighting the restrictions (a fight the national-security framing makes very hard to win) and toward accelerating allied coordination — pushing Washington to close the Japan/Netherlands enforcement lag that is currently rewarding Tokyo Electron at Lam Research's and Applied Materials' expense.
  • Treat the U.S./allied hyperscaler boom as the real strategy, not the consolation prize — China's share of Nvidia's datacenter revenue was already down to roughly 9%; formalizing that shift rather than fighting it converts an externally imposed constraint into a deliberate portfolio decision.
Trade-off I'm not glossing over: this path effectively cedes the long-run Chinese AI-compute market to Huawei, Biren, and Moore Threads. Given the Kirin 9000S and Ascend 910B evidence in Block 4, that could produce a formidable, fully independent Chinese compute ecosystem within the decade. But the alternative — continuing to chase shrinking legal headroom inside China — risks worse outcomes: further-restrictive follow-on rules, reputational damage with the Commerce Department, and engineering cycles spent on chips that get banned before they ship. Between a bounded, known cost and an open-ended regulatory chase, the bounded cost is the more defensible bet.
Block 6 — Official Case Discussion Questions
Q1. When are export controls a good idea? Under what conditions are they likely to succeed?
Export controls succeed when four conditions hold together: (1) the technology is a narrow, identifiable chokepoint rather than a diffuse capability; (2) the regulator has real enforcement capacity to verify end use and close loopholes; (3) every allied holder of a related chokepoint (fabrication, lithography, tools) restricts in coordinated lockstep, not years apart; and (4) the target cannot quickly substitute a domestic alternative. The case's own history — Cold War cryptography controls being abandoned as "futile" once the U.S. recognized it couldn't control the diffusion of mathematics, and pivoting instead to leading and setting global standards — is the clearest illustration that controls buy time, not permanent containment. See Block 3 for the full argument, including the Diplomat's account of how uncoordinated timing let Japanese and Dutch firms capture revenue Washington intended to deny China.
Q2. Are advanced semiconductors a good candidate for export controls? Why or why not?
Yes, more so than most technologies — CSIS's "four chokeholds" (chip exports, EDA software, manufacturing equipment, U.S.-built components) are genuinely narrow and hard to replicate. But applying the same four conditions honestly: enforcement has visible gaps (CSIS's own follow-up report called the picture "mixed"), allied coordination arrived late enough that China stockpiled over $40 billion in equipment in 2023 alone, and China's substitution capacity is moving faster than initial 2022 assessments assumed — Huawei's Kirin 9000S shipped just 11 months after the first restrictions, without EUV lithography or U.S. EDA tools. My position: advanced semiconductors clear the technology-specificity bar well, but the October 2023 execution outran the enforcement and coordination conditions needed for controls to work as intended. See Block 4 for the full breakdown against each condition.
Q3. How should Nvidia respond to the U.S. government's attempt to restrict the export of advanced semiconductors?
Stop treating chip redesign as a durable strategy — Secretary Raimondo's explicit warning and the shrinking performance ceiling on compliant chips (down to roughly 2017's V100 level) make the cat-and-mouse approach a dead end. Instead: comply substantively rather than technically to avoid further enforcement actions; make DGX Cloud services the primary vehicle for any remaining legal China exposure, since it preserves revenue with better end-use visibility; redirect lobbying away from fighting the restrictions (a losing battle given the national-security framing) and toward accelerating allied enforcement coordination; and treat the U.S./allied hyperscaler demand boom — not China — as the strategic center of gravity going forward. This cedes long-run ground in China's AI-compute buildout to Huawei and other state-backed rivals, a real and acknowledged cost, but it converts an open-ended regulatory chase into a bounded, manageable one. See Block 5 for the full options analysis and trade-offs.
Block 7 — Participation Hooks

Consensus Point

The class will quickly agree that Nvidia has no choice but to comply with U.S. national-security rules — that's table stakes, not a strategic decision. Don't spend airtime relitigating whether Nvidia should defy Washington; move straight to how it should compete within the constraint.

Provocative Push

Push the room on this: is the export-control regime manufacturing its own failure condition? A Chinese chip industry that ships a competitive domestic AI accelerator (Kirin 9000S) 11 months after restrictions begin, and whose own CEO claims parity with GPT-4, suggests the controls may be accelerating exactly the self-sufficient Chinese tech ecosystem they were designed to prevent — at the cost of near-term U.S. firm revenue and global market share along the way.

Taju's Edge — A Bifurcated Compute Ecosystem, Seen From an AI PM's Desk

As a product manager building AI infrastructure at Redamo Labs, my sharpest lens here isn't the U.S.-China axis — it's what a bifurcating compute stack (CUDA/Nvidia vs. a Huawei/domestic Chinese alternative) means for AI builders outside both blocs. African and emerging-market startups don't manufacture chips or write export policy; they consume compute through hyperscalers built on Nvidia GPUs, and every dollar of scarcity or price premium those restrictions create flows straight through to a Lagos or Nairobi startup's AWS bill. This case isn't just about two great powers — it's a live cost-of-compute-access question for every builder in the "third bloc."

Taju's Edge — Standards Fragmentation as a Non-Alignment Problem

If the compute world splits into a CUDA-based stack and a Huawei/domestic Chinese stack, emerging-market developers face a version of the Cold War non-aligned movement's dilemma: build for one ecosystem and risk lock-in, or build for both and pay a real engineering tax. Worth raising as a live example of how great-power technology competition exports costs to bystander markets — exactly the systems-level "who wins, who loses" thinking the Group Strategic Risk Assessment rubric asks for.

Block 8 — Capstone: Where This Connects Across the Whole Course

As the final session, this case is less a new topic than the place where the course's threads converge — the firm-level answer to six sessions of country- and industry-level analysis.

Session 5 — Taiwan & Semiconductors

This session is that one's export-control theme playing out inside a single company. TSMC's fabrication chokepoint and the "silicon shield" debate are the upstream reality Nvidia depends on and cannot control — every strategic option in Block 5 assumes Nvidia can still get chips manufactured at all.

Session 1 — Singapore

Singapore's "Global-Asia," proactive-neutrality positioning is itself a version of the political-engagement and regulatory-navigation strategy studied here — a small, trade-dependent actor threading U.S. security ties and Chinese economic ties, the same balancing act Nvidia performs at the firm level instead of the state level.

Group Assignment Template

This session's framework — what structural change is occurring, why it matters strategically, who wins and loses, what management should do — is literally the Strategic Risk Assessment rubric. Nvidia is a ready-made template for how to structure that deliverable end to end.