Who joins, who pays, and which innovation deserves a separate experiment?
Session 7 connected a strategic position to new demand and the capabilities that sustain it. Here, the two assigned readings challenge how a firm develops that position: platform strategy asks whose participation creates value; disruption theory asks whose needs an established business may overlook. Use both lenses to interrogate NVIDIA’s 2024 case.
NVIDIA sold graphics cards to gamers. In 2006 it launched CUDA, a way to program those chips for other work, and for
years few people needed it. By the second quarter of 2024 data centres brought in almost 90% of its revenue, and
analysts put its 2023 share of the data-centre market at 98% (case, p. 6). Which part of this rise depends on technology, complementary applications, or a change in the customers being served?
01When does a new technology create a new market, and when does it only sharpen competition in an old one?
02How did NVIDIA turn a graphics chip into a platform, and which part of its advantage is not the chip?
03Which innovation is sustaining, which could be disruptive, and what evidence would settle the distinction?
NVIDIA facts on this page come from the Yoffie and von Bargen case, set in late 2024, with page numbers. Where the
page reads the case through a framework, it says so. Nothing here reports events after the case was written.
NVIDIA returns in Session 9
for its corporate scope.
Session 8 · your learning targets
Learning outcomes
By the end of this session, you should be able to:
Map distinct participant groups and explain a cross-side benefit, a same-side effect and the platform rules connecting them. Map the participation mechanism →
Defend a pricing or subsidy hypothesis using price sensitivity, complement quality, multi-homing and evidence of retained use. Test the platform assumptions →
Distinguish sustaining improvement from a potential disruptive trajectory by identifying the initial customer, performance trade-off and later improvement path. Classify the innovation →
Explain how CUDA and complementary applications may reinforce NVIDIA’s hardware offering while distinguishing ecosystem analysis from evidence of market tipping. Trace the CUDA mechanism →
Recommend a bounded innovation experiment using the 2024 NVIDIA case, with a measurable learning milestone and a condition that would reverse the recommendation. Defend a decision →
Use the readings, case evidence and practice to explain your reasoning. Browser checks provide practice feedback; formal assessment follows the course outline and instructor guidance.
Recall → connect → carry forward · Sessions 1–7
The strategy story so far: Sessions 1–7
Sessions 1-7 move from strategic choice to a defensible claim about relative value and cost. Each step constrains the next: ambition must survive industry economics, ecosystem responses, capability limits and the threat of imitation. Nila Foods is a fictional Indian meal-service business used to make that chain visible. Its possible gains remain hypotheses. Today asks how innovation and complementary participation might strengthen its focused offer, and what would count as evidence.
One continuing classroom story. Nila Foods is a fictional Indian meal-service for working adults. Follow its choices from Session 1; later expansion ideas are proposals to test.
Operational effectiveness improves similar activities; strategy requires positioning, trade-offs and fit. State objective, scope and advantage, then compare intended plans with emergent action using evidence.
Nila Foods: Nila proposes preordered weekday meals on clustered routes, declines instant citywide delivery and studies actual customer behaviour.
Purpose and values guide what endures, while a quantified ambition opens a same-date gap from the baseline; closing that gap requires external and internal evidence.
Nila Foods: Nila separated easing meal burdens and honest food-safety disclosure from its growth target, then compared that target with its projected baseline.
Industry boundaries, PESTEL and five forces connect external change to price, cost and investment; complements and possible responses then raise questions about the wider game.
Nila Foods: Nila traced ingredient-price pressure, supplier bargaining, lunch substitutes and easy entry before asking whether a neighbouring seller could also help it.
Value nets and PARTS show how co-opetition changes added value, while uncertainty shapes commitments; credible moves therefore require examining internal strengths.
Nila Foods: Nila considered sharing a collection point with a rival and testing the rules through a limited pilot while demand remained uncertain.
Resources and capabilities support advantage when they improve buyer benefits or relative costs and pass tests against rivals and substitutes; value-chain analysis then checks their economic contribution.
Nila Foods: Nila examined kitchen routines, route know-how and trust, asking which linked activities rivals could imitate and whether it could retain their benefits.
Use the value stick to distinguish value created from value captured. Test shared value against both business and social effects, including burdens shifted elsewhere.
Nila Foods: Nila asks whether demand forecasts and reusable containers reduce waste and cost, including the return and cleaning work.
Porter distinguishes cost leadership, differentiation and focus; Kim and Mauborgne test value innovation; Ghemawat and Pisano examine threats, commitments and capabilities.
Nila Foods: Nila compares dependable delivery and relative cost within its chosen segment. Its value curve and four-actions proposal must demonstrate new demand and viable economics before supporting a blue ocean claim.
Why this leads to Session 8
Suppose Nila invites nutrition creators to complement its weekday meals. A list of creators does not establish a platform: identify distinct sides, explain why an additional participant benefits the other side, and test price sensitivity, quality, same-side effects and multi-homing. Keep the experiment tied to its chosen customers and cost position. Separately, test any disruption claim against an initial foothold, a performance trade-off and a path toward mainstream needs.
Retrieve it before moving on
Reconstruct the seven-session chain behind Nila's creator proposal. What would make the proposal a useful complement to its strategy, what additional evidence would support a two-sided platform explanation, and why would neither establish disruption?
Compare your reasoning with a model answer
Nila begins with a focused weekday promise and reinforcing preorders and routes. Purpose and a dated value gap explain what improvement it seeks. Five forces test the economics; co-opetition and uncertainty test how office partners might respond and what a pilot could reveal. The resource audit asks whether its coordination routines outperform alternatives. The value stick and social ledger test whether reduced waste creates economic and social gains. Session 7 then asks whether those gains support differentiation, relative low cost or a credible value-innovation proposal. Creator content should serve that position rather than widen the menu indiscriminately. A platform explanation additionally needs evidence that customers attract useful creators and creators improve retained customer use, under workable prices and rules. Disruption requires a distinct foothold and improvement trajectory relative to an incumbent; novelty or growth alone is insufficient.
Read the argument · compare the lenses
Platforms and disruption: two different explanations
Who must participate for whom to benefit?
A two-sided platform connects distinct participant groups whose demand is interdependent. Map the sides before designing prices: an extra user can raise value for the other side, while extra participants on the same side may bring learning or congestion. The price structure matters as well as the total price.
Use three reading questions: which side should be subsidized; when might a market tip toward one platform; and how could an adjacent platform bundle its way into your market? Check price sensitivity, cross-side benefits, multi-homing and differentiated needs before assuming a winner-take-all outcome. These are hypotheses to investigate, not automatic properties of every software product.
Case application: Draw a candidate CUDA ecosystem map linking developers, applications and GPU users. Then explain the transaction or complementary benefit connecting each pair. A software toolkit is not automatically a marketplace; specify the mechanism you mean.
Can listening to your best customers hide the next market?
The assigned 1995 article asks why established firms can miss emerging technologies. Read it through the performance demanded by current customers and the investment filters serving those customers. A technology can be initially unattractive on mainstream measures yet useful in another setting.
The later theory distinguishes a foothold among overserved buyers or nonconsumers and subsequent improvement toward mainstream needs. That trajectory matters: a faster premium product may be sustaining innovation, even when technically radical. An autonomous experiment can use different customers, milestones and economics when the existing business would screen it out.
Case application: Do not label NVIDIA “disruptive” solely because it became successful. Name the incumbent, initial customer, performance disadvantage, different benefit and improvement path. Mark any missing evidence as missing.
Core question · the platform business model
What makes participation on one side worth subsidizing?
Build the causal map
For a hypothetical AI toolkit: more useful applications could attract more users; more users could make development worthwhile. These two arrows are cross-side hypotheses. More developers could also intensify competition among developers. Keep that same-side effect separate.
Choose a measurable bottleneck
A small network may need complements before users arrive, but free access alone may not solve poor quality or a weak use case. Track useful active complements, user retention and willingness to pay. Downloads are not a sufficient test of network value.
NVIDIA evidence task: Use the case’s CUDA history and customer discussion to identify which participants supply complements, who buys hardware, and who could use multiple ecosystems. Explain where the two-sided lens is helpful and where NVIDIA’s vertically connected product system requires another explanation.
Session 7 closed with
Porter’s three conditions under which a firm can be both low-cost and differentiated, each of them temporary. The
third is today’s starting point: a firm that pioneers a major innovation can lower cost and raise quality together,
but only while nobody else has the innovation.
Session 7 · Porter (1985)
Choose, or be stuck in the middle
Uniqueness usually costs money, so cost and differentiation pull
against each other, and the firm that will not choose gets neither. Kim and Mauborgne attack the premise: the trade-off
binds firms that compete on the factors their industry already competes on, and breaks for firms that redraw them
(2004, pp. 14–15).
Session 7 · Porter’s third condition
Innovation buys a temporary exemption
“The ability to be both low
cost and differentiated is a function of being the only firm with the new innovation.”
Porter (1985, p. 20)
Keep both readings open: is NVIDIA in a new market space with its
own barriers, or in Porter’s temporary exemption?
In
Sustaining advantage extension the
barriers were commitments and capabilities. Kim and Mauborgne give blue oceans their own list: scale economies, network
externalities, whole-system change and brand (2004, pp. 19–21). CUDA will test the second.
The trap this session sets
It is easy to call any spectacularly profitable new technology a blue ocean. Kim and Mauborgne say nearly the opposite:
across a century of examples, blue oceans were seldom the product of technological innovation as such. They came from
linking technology to what buyers valued, often by simplifying it (2004, pp. 11–12). Hold NVIDIA to that
test. Was its blue ocean the GPU, or what CUDA let buyers do with the GPU?
Kim & Mauborgne (2004) · Blue ocean strategy
Transfer from Session 7: can NVIDIA’s value curve explain its growth?
Kim and Mauborgne split the business universe in two. Red oceans are the industries that exist today: boundaries are
accepted, the rules are known, and firms fight for a bigger share of existing demand until products turn into
commodities (pp. 2–3). Blue oceans are the industries not yet in existence, where demand is created rather
than fought over (p. 3). A few are wholly new industries, as eBay’s online auctions were. Most are carved out
of a red ocean when a firm moves an industry boundary, as Cirque du Soleil did between circus and theatre (p. 3).
Two logics of strategy
Rewritten in this page’s words from Kim & Mauborgne (2004, p. 14).
Red ocean
Blue ocean
Fight inside the market space that already exists
Open market space that nobody contests
Set out to beat rivals
Make rivals irrelevant
Win a larger share of today’s demand
Create demand and capture it
Accept that more value costs more
Break the value–cost trade-off
Line up every activity behind differentiation or low cost
Line up every activity behind differentiation and low cost
The evidence they offer
Business launches at 108 companies (Kim & Mauborgne, 2004, p. 6).
Line extensions86% of launches
Revenue62%
Profit39%
New markets or industries14% of launches
Revenue38%
Profit61%
Read it carefully: the authors decided which launches counted as
market-creating, and the article does not say how many such launches lost money.
The defining move
Value innovation
Conventional strategy says a firm can offer more value at higher
cost or reasonable value at lower cost. Blue-ocean creators reject that tenet and pursue both at once (pp. 14–15).
Buyers get value from utility and price; the firm gets value from price and its cost structure, so the leap happens
only when utility, price and cost are aligned as one system (p. 18).
“Value innovation is the
simultaneous pursuit of radically superior value for buyers and lower costs for companies.”
Kim & Mauborgne (2017, p. 21; first published 1997)
Worked example · pp. 21–22
Ford’s Model T priced against the horse
Over 500 American carmakers built handmade luxury cars at about
$1,500. Ford looked outside the industry for its price point, at the $400 horse-drawn carriage, and built a car for
everyday use: one colour, few options, easy to drive and fix. Standard parts and the assembly line cut build time to
four days against the industry’s 21.
$850 → $290
price, 1908 to 1924
9% → 61%
share, 1908 to 1921
4 days
per car, vs 21
Four findings that matter for a technology company
pp. 11–12
Not about technology as such
The underlying technology usually existed already; Ford’s line came from meatpacking. In computers, blue oceans came from linking technology to what buyers valued, often by simplifying it.
p. 12
Incumbents make them too
GM, Chrysler, IBM and Compaq were established players when they opened theirs, usually inside their core business. Blue oceans are often right next to you.
pp. 12–13
Study the move, not the firm
The same company can be brilliant at one time and wrong-headed at another. Compaq vanished into HP in 2001, yet its PC-server move created a multibillion-dollar market.
p. 19
Structure is not destiny
Red-ocean thinking treats industry structure as given, the structuralist view. Blue-ocean thinking assumes players can redraw it, the reconstructionist view.
Why blue oceans stay blue for a while
Creators usually go 10 to 15 years without a credible challenge,
as Cirque, Home Depot, FedEx, Southwest and CNN did (p. 19). The barriers are economic and cognitive (pp. 19–21):
scale that new volume brings fast (Walmart’s purchasing); network externalities,
where more users make the offer more attractive (eBay); whole-system change that rivals’ internal
politics resist (airlines copying Southwest); and brand, from the loyalty a leap in value earns
(Intuit’s Quicken held off Microsoft for over ten years) to imitations that would contradict the copier’s own
image (The Body Shop). Network externalities are the barrier to watch when we reach CUDA.
The tools · Kim & Mauborgne (2004, 2017)
The strategy canvas and the four actions
The 2004 article names the four moves behind value innovation: eliminate and reduce to cut cost, raise and create to lift
buyer value (pp. 18–19). The tools that put those moves on one page come from two earlier articles by the same
authors, reprinted in their Blue Ocean Strategy Reader (2017): the value curve and the four questions
(pp. 10–12), and the strategy canvas (pp. 93–96).
Tool 1 · Reader, pp. 94–96
The strategy canvas
List the factors an industry competes on along the horizontal axis.
On the vertical axis, mark how much each player offers on each factor, or, for price, how much it asks. Join the dots
and you have each player’s value curve. One picture shows the industry’s factors, the rivals’ profiles
and your own.
Three tests of a good curve
1Focus. Southwest stresses three factors (friendly service, speed and frequent point-to-point flights) and prices against the car, not other airlines.
2Divergence. Reactive rivals share one curve; Southwest’s competitors were so alike that the authors drew them as one line.
3A tag line that is true. “The speed of the plane at the price of the car—whenever you need it” (p. 96). If you cannot write one, you probably lack a strategy.
Tool 2 · Reader, pp. 11 and 21
The four actions
Eliminate · cuts cost
Which factors the industry takes for granted should go? Tests whether they still deliver any value.
Reduce · cuts cost
Which should fall well below the industry standard? Finds what was over-designed in the race to match rivals.
Raise · lifts value
Which should rise well above the standard? Removes compromises the industry forces on buyers.
Create · lifts value
Which should exist that the industry has never offered? Breaks the industry’s boundaries.
Asking all four, not one or two, is what makes growth profitable
(p. 21). A curve that keeps the industry’s shape at a slightly different level is still a red-ocean curve:
rivals offering “a little more for a little less” rarely challenge the shape (p. 12).
Cirque du Soleil through the four actions
Assembled from the article’s text, Kim & Mauborgne (2004, pp. 15–17)
Eliminate
Animal acts, costly to keep and losing public favour
Star performers
Three-ring shows
Aisle concession sales
Reduce
The role of acrobats and thrill acts: kept, but shorter and more artistic
Raise
The tent: a glamorous exterior and real audience comfort, no sawdust or hard benches
Clowns kept, their humour made more sophisticated
Create
A theme and storyline
Several productions, so people come back
An original score that drives the show
Abstract dance borrowed from theatre and ballet
The result pulled in adults and corporate clients who had been
going to the theatre and would pay several times a circus ticket (p. 2), while dropping the most expensive parts of
a circus (p. 17). That is the test for every curve on this page: new buyers, and a lower cost base, at once.
When it goes wrong · Rothaermel (2024, p. 234)
A blue ocean attempt that fails to reconcile value and cost leaves
the firm with neither a clear differentiation nor a clear cost profile: Session 7’s stuck in the middle.
Rothaermel’s example is JCPenney, which tried to move upmarket and sailed into a red ocean instead (pp. 234–235).
After the leap · Reader, pp. 14–15
Imitators arrive eventually. Compaq’s first PC was a new value
curve; then it spent years trying to beat IBM feature for feature, and by the late 1980s the two curves were nearly
identical. The authors’ advice is to watch for converging curves and value-innovate again. Keep Compaq in mind
for NVIDIA’s product cadence.
Innovation as strategy formulation · Rothaermel (2024), Grant (2019)
Types of innovation: what is new, the technology or the market?
A blue ocean describes a market outcome. To choose a technology strategy you also need to say what kind of innovation you
are making. Rothaermel sorts innovations with two questions: does it build on existing technology or a new one, and does
it target an existing market or a new one (pp. 274–275)? Four types follow, in two diagonal pairs.
Market
New market · existing technology
Architectural
Known components reconfigured to attack a new market. Canon redesigned the copier so small offices could run it without a service contract, a segment Xerox ignored (pp. 276–277).
New market · new technology
Radical
Novel methods or materials, often a new knowledge base. The jet airliner: de Havilland commercialised it first, Boeing rode it to dominance (p. 275).
Existing market · existing technology
Incremental
Steady improvement on an established knowledge base; most innovation is of this kind. Gillette added blades, one at a time, from one to six (p. 275).
Existing market · new technology
Disruptive
Starts as a cheaper, initially inferior answer to an existing need, improves faster than buyers need, and invades from the bottom up. Netflix streaming against the cable bundle (pp. 277–278).
Technology → existing · new
Rothaermel, pp. 275–276
Why leaders drift to incremental
Three forces: economic incentives (a leader
protects the position radical change would threaten), organisational inertia (processes and power
built around the current product) and the innovation ecosystem (suppliers, buyers and complementors
that incremental change keeps happy and radical change disrupts). New entrants carry none of this.
Grant, pp. 222–225
Creating value is not capturing it
An innovation pays the innovator only if it keeps a share of the
value it creates, its regime of appropriability. Four things decide that share: property rights,
the tacitness and complexity of the technology, lead time, and control of complementary resources. Grant names
Nvidia in graphics chips among the firms “brilliant at exploiting lead time” (p. 225).
Activity · classify the innovation
Ten innovations, four types
Done
Six NVIDIA-era moves from the case and four textbook
examples. Pick the type before you read the explanation. Where a second answer is defensible, the page says so and
shows it in amber: the argument matters more than the label.
0 of 10 classified.
The case · Yoffie & von Bargen (2024), pp. 1–6
Application of technology: from graphics chips to AI
On 18 June 2024 NVIDIA was briefly the world’s most valuable company, worth $3.4 trillion (p. 1). Its revenue
had grown from $159 million in FY2000 to $60.9 billion in FY2024 (p. 2). The route was not planned in
advance. Read it as one technology, parallel processing, looking for the applications that would pay for it.
1993 · p. 2
Gaming as the killer app
Games were among the hardest computing problems and sold in huge volume. Huang: “‘Video games’ was our killer app.”
1996–99 · p. 3
RIVA 128, then the GPU
With nine months of cash, NVIDIA cut an 18–24-month chip cycle to under six using software prototypes and sold a million RIVA 128s in four months. GeForce 256 (1999) was marketed as the first GPU.
2000s · p. 3
Parallel, and fabless
GPUs run many computations at once where CPUs work sequentially. NVIDIA designed the chips and left fabrication mostly to TSMC, escaping the capital a fab needs: up to $25 billion for one by 2024.
2006 · p. 4
CUDA
A C++-like language that became a platform for programming GPUs beyond graphics. Proprietary to NVIDIA hardware; free development kits for everyone.
2010–13 · p. 5
The lean years
CUDA downloads fell and, with the stock depressed, the board worried about raiders. Machine learning rescued CUDA.
2012–13 · p. 5
AlexNet and the bet
Three academics broke the ImageNet record using CUDA on NVIDIA GPUs. Huang then bet on deep learning while AI revenue was small: “we were no longer a graphics company,” one VP recalled.
2017–22 · p. 5
The crypto bridge
Miners bought GPUs by the millions; NVIDIA launched a mining line in 2017 to protect gaming supply. Mining revenue halved in the Q4 2018 crypto winter.
2022–24 · p. 6
Generative AI
ChatGPT launched in November 2022. OpenAI said training GPT-4 took 25,000 NVIDIA GPUs and $100 million; NVIDIA’s were the only chips able to train such models, and most developers used CUDA.
Exhibit workshop · case Exhibit 1a, p. 14
The revenue mix tells you when the bet paid
Done
Data centreGamingEverything else
62% → 73%
gross margin, FY2020 to FY2024 (Exh. 1a)
$3,100
estimated cost of an H100 sold for $25,000–40,000 (Exh. 3)
$8.6B
R&D in 2023, up 17% but only 14% of sales vs AMD’s 20% (p. 6)
~3.7M
GPUs sold in 2023 and a 98% data-centre share, analysts’ estimates (p. 6)
Exhibit 1a, $ millions, fiscal years ending January. “Everything else” is revenue minus the two named segments, computed here.
Fiscal year
Revenue
Data centre
Gaming
Everything else
Data centre share
What the path teaches about applying technology
Hardware arrives first
Chip performance often arrives before the applications that can use it, and NVIDIA’s chips were among the most expensive to make (case, p. 4). CUDA was Huang’s answer: hand developers the tools to invent the uses. The technology choice was a market-creation choice.
Strategic patience
The professor’s teaching script calls innovation a marathon: invest in foundations before the killer app is obvious. CUDA had falling downloads for three years, and customers adopted deep learning slowly from 2012 to 2020 (p. 5). The bet survived on gaming and crypto revenue.
Serendipity, then commitment
AlexNet came from academics and crypto from miners; neither was NVIDIA’s plan. What NVIDIA chose was to commit once the signal came, which Session 1’s Mintzberg would call an emergent strategy made deliberate.
Where NVIDIA should sit in the value chain (complete DGX
systems, the DGX Cloud rental business, the Mellanox purchase, the abandoned Arm deal and its dependence on TSMC) is a
question of corporate scope, and it belongs to
Session 9.
Platforms and ecosystems · case pp. 4, 11–12 · Rothaermel, Grant
CUDA: how a chip became a platform
A chip on its own is a pipeline product: design it, have it made, sell it. A platform, in Rothaermel’s definition,
creates value by matching external producers and consumers on infrastructure the platform owner manages
(p. 282). CUDA moved NVIDIA from the first to the second. It began in 2006 as a programming language, then NVIDIA let
outside developers write programs for it and share them with other developers, with the development kits free
(case, p. 4). By 2022 Huang counted over 150 kits serving fields from gaming and design to life sciences, quantum
computing, AI, cybersecurity, 5G and robotics (p. 4).
The CUDA ecosystem, mapped onto Rothaermel’s four roles (p. 283)
Owner
NVIDIA. CUDA is proprietary: software written in it runs only on NVIDIA GPUs (case, p. 4).
Producers
Developers and researchers who write and share CUDA programs, like the AlexNet team (pp. 4–5).
Consumers
Buyers of accelerated computing: gamers, scientists, miners, then AI labs and data centres (pp. 3–6).
Providers
The cloud companies that rent NVIDIA GPUs to others, profitably, while designing their own chips (p. 1).
The professor’s flywheel, checked against the case
1GPUs process in parallel, which suits graphics and, it turned out, neural networks (pp. 3, 5).
2CUDA and its free kits make that power usable by developers (p. 4).
3Shared programs and libraries give the next developer more reason to build on NVIDIA (p. 4).
4Demand and premium prices fund R&D for the next GPU: $8.6 billion in 2023 (p. 6). Back to 1.
The teaching script calls this a walled garden. The case gives no
count of CUDA developers, so this page does not either.
Grant, p. 234
Network externalities
A product is worth more to each user as more people use it. Two
sources fit CUDA: complementary products (the libraries others write) and saved switching costs (code and skills that
already work). Positive feedback can tip a market to one winner, and collective lock-in keeps a standard in place
even after better ones appear.
Grant, p. 236
Open to grow, closed to earn
Platform owners trade market penetration against appropriation.
JVC licensed VHS and beat Sony’s closed Betamax; IBM’s open PC won the market while Intel and Microsoft
took the value. NVIDIA’s answer was to give the software away and lock it to the hardware it sells.
Kim & Mauborgne, p. 20
The eBay barrier
The more customers eBay had, the more attractive it became to buyers
and sellers alike, leaving users little reason to go elsewhere. That is the blue-ocean barrier CUDA most resembles,
and the one rivals in 2024 were spending most to break.
Moat or swamp? The 2024 attack on CUDA
The moat
Customers worried that code written in CUDA could not move to other
chips. An AWS vice-president put the problem plainly: “Nvidia’s got great chips, and more importantly, they
have an incredible ecosystem,” which made it very hard to get customers onto other chips (p. 11). AMD’s
open alternative, ROCm, launched in 2016, never matched CUDA’s popularity (pp. 11–12).
The swamp
OpenAI’s open-source Triton (2021) lets researchers without
CUDA experience write efficient GPU code that runs on many makers’ chips; by 2024 Microsoft, Meta and Google were
investing in it. Intel’s CTO: “Essentially it breaks the CUDA lock-in.” Tenstorrent’s Jim
Keller called CUDA “a swamp, not a moat,” too complicated, and predicted a UNIX-to-Linux ending (p. 11).
Notice who is backing Triton: three of NVIDIA’s largest
customers. A platform’s complementors and buyers can also be its most motivated attackers, which is
Rothaermel’s innovation-ecosystem point seen from the other side.
The case · demand and competition, pp. 4–12
AI demand, rivals and custom chips
Kim and Mauborgne expect a blue-ocean creator to go 10 to 15 years before a credible challenge (2004, p. 19). CUDA
launched in 2006 (case, p. 4). By 2024 the challengers had arrived, and they were not all attacking the same thing.
Demand: how big, and for what?
Size. Some analysts saw a $1 trillion opportunity for chip makers over a decade; others feared a telecom-style overbuild, with no killer app yet and a query costing up to seven times a Google search (pp. 7–8).
Training or inference. GPUs were essential for training, with no good alternative in 2024. For inference there were several, often with lower total cost of ownership (p. 8).
Smaller models. Meta’s open-source Llama came in sizes from 8 billion parameters; Apple ran a small model on the iPhone and trained on Google’s TPUs, not NVIDIA’s GPUs (p. 8).
Power. One NVIDIA GPU could draw 1,000 watts or more, and analysts called power “the ultimate constraint” (p. 4).
Direct rivals: AMD and Intel
AMD. Its MI250X delivered about 380 teraflops against up to 1,000 for the H100, and sold for $8,000–10,000 against $25,000–40,000. It raised its 2024 data-centre GPU forecast from $4.0 to $4.5 billion but could get only a few hundred thousand GPUs from TSMC (p. 9).
Intel. Habana’s Gaudi2, bought for $2 billion in 2019, was priced near half an H100, but trailed on performance and its software was less mature than CUDA (p. 9).
Start-ups. Cerebras claimed inference 20 times faster at a fifth of the cost per token; Groq’s chips were fast but specialised. Venture capital put $6 billion into AI chip firms in 2023 (pp. 10–11).
Customer-competitors
Google. TPUs are chips designed specifically for neural networks. TPU v4 drew 200–250 watts against 250–400 for NVIDIA’s V100 and A100, and cost Google $2,000–3,000 each against about $15,000 for an NVIDIA chip (p. 10).
Amazon. Trainium trains models for firms such as Anthropic; Inferentia runs inference for Airbnb and ByteDance. Amazon claims up to 40–50% better price-performance and wants to cut the “Nvidia tax” (p. 10).
Concentration. Nearly half of NVIDIA’s revenue came from four large data-centre customers (p. 6).
Activity · strategy-canvas builder
Draw the challenger’s value curve against NVIDIA’s
Done
NVIDIA’s 2024 curve is fixed. Load a challenger, then
move its scores with the − and + buttons, or add a factor NVIDIA does not offer. The four-actions grid and the verdict
update as you go. Scores are this page’s reading of the case on a 0–10 scale, not case data; each row shows its
evidence. For price, higher means the buyer pays more.
NVIDIA, 2024Challenger
Eliminate
Reduce
Raise
Create
The verdict · Kim & Mauborgne against the case
Did NVIDIA create a blue ocean, and is it turning red?
Apply the reading’s tests, not the share price. A blue ocean creates new demand, makes rivals irrelevant and breaks
the value–cost trade-off for buyers and the firm together. The case supports both sides.
The case for a blue ocean
New buyers, not stolen ones. Scientists in the biosciences, computing and chemistry found NVIDIA GPUs did calculations in a fraction of the time of standard scientific computers (case, p. 5). They were not graphics customers. Cirque, too, won people who had never bought circus tickets (2004, p. 2).
Technology linked to value. Parallel processing already existed in gaming GPUs (p. 3). CUDA was what let buyers use it for their own problems (p. 4): value pioneering in the authors’ sense, not technology for its own sake (2004, pp. 11–12).
Uncontested for years. Rivals were slow to follow: Intel cancelled its high-end graphics chip after four years of work, and AMD’s answer to CUDA arrived only in 2016 (pp. 9, 11). In 2023 NVIDIA held an estimated 98% of the data-centre market (p. 6).
The case against
No low-cost half. Buyers got a leap in performance at “outsized premiums” (p. 1): an H100 estimated to cost $3,100 sold for $25,000–40,000 (Exhibit 3). Customers call the difference the “Nvidia tax” (p. 10). The Model T took share by pricing against the horse.
Porter’s exemption, not a new logic. Being low-cost and differentiated at once lasts only while a firm is the sole owner of the innovation (Porter, 1985, p. 20). By 2024 rivals had chips, and Triton was going after the software (pp. 9–11).
Racing on the old factors. NVIDIA was moving to annual launches; Blackwell promised 2.5 times the performance for 25% more cost (p. 6). That is Compaq racing IBM on features, the convergence trap (Reader, pp. 14–15).
Activity · red or blue?
Eight moves from the case. Which ones create demand?
Done
For each move, decide whether it creates new demand (blue)
or fights for demand that already exists (red). The test is Kim and Mauborgne’s, not whether the move is clever or
profitable. When you have judged all eight, the page sums up what you have said about NVIDIA’s ocean.
0 of 8 judged.
Where this leaves NVIDIA
The fairest reading of the case is that NVIDIA opened a blue ocean
with CUDA and the deep-learning bet, priced it as the only supplier, and by 2024 was defending it with red-ocean
weapons: speed, performance per dollar and lock-in. The authors would say the signal to value-innovate again is a curve
that rivals are starting to copy. “Value innovation is about offering unprecedented value, not technology or
competencies” (Kim & Mauborgne, 2017, p. 15). What the next move should be, and how far up the value
chain NVIDIA should go to find it, is where
Session 9 begins.
Go deeper
Five earlier decks on the same ground
Each was built for another course or an earlier numbering, and the
NVIDIA decks draw on sources outside this session’s case. Where their figures differ from this page, this page has
been checked against the case and the reading. Use the decks for practice and simulation.
A small, awkward market asks for a simpler product
Fictional teaching scenario: your AI tool serves large enterprise buyers. A small team proposes a cheaper tool for organizations that cannot use your flagship product. It performs worse on the flagship benchmark. What would you fund?
Consider the consequences
Their feedback is essential for the current business, but they may correctly reject a product intended for someone else. Interview the proposed users directly. Ask what they do today, what is inaccessible, and which performance measures matter to them.
Consider the consequences
Give the team a defined learning budget and milestones suited to the proposed customers: actual use, repeat demand, delivery cost and improvement potential. Independence is useful only if the opportunity needs a different business model; it is not a licence to ignore evidence.
Consider the consequences
A label does not demonstrate a foothold or an improvement path. A small-market hypothesis can be wrong. First test whether the new customers value the different benefit and whether the economics can work at their price.
Platform studio · 12 minutes
Who pays to get the network started?
Design a fictional AI-tool platform linking developers and enterprise users. Change developer fees and cross-side value. Every number is illustrative; this is a one-period sensitivity model, not a forecast of NVIDIA or a calibrated demand model.
Model: developers = 100 − 8 × developer fee. Enterprise users = 20 + cross-side benefit × developers. Enterprise fee = 8; service cost per enterprise user = 3; fixed cost = 100. Fractional user counts represent an expected demand index.
Developers
Enterprise users
Platform surplus after fixed cost
Surplus at zero developer fee
This model includes one cross-side link. It omits reverse feedback, enterprise price sensitivity, quality, competition and multi-homing. A positive result is conditional on the assumptions.
Your one-minute recommendation
Write a claim, cite one case exhibit or reading page, and name evidence that would change your mind. This note stays on the page until you leave; download it to keep a copy.
For class
Six questions to argue
Come with a position on each and the page of the case or reading
that supports it.
01
Map the participant groups around CUDA. Which cross-side benefit is supported by the case, and which remains a hypothesis? Identify one same-side effect separately.
02
A free developer toolkit can support paid hardware sales. Apply Eisenmann, Parker and Van Alstyne: which participation constraint does the subsidy address, and when might it fail?
03
Could users and developers participate in multiple ecosystems? Use one case fact to challenge the claim that network effects must produce a single winner.
04
Use Bower and Christensen to diagnose a candidate innovation in the case. Identify its first customers, its initial performance disadvantage and its improvement path. If the evidence is missing, say so.
05
Which idea would NVIDIA’s most demanding customers reject even if another group valued it? Design a bounded entrepreneurial experiment with a customer, a learning milestone and a stop condition.
06
Compare a value-innovation explanation from Session 7 with the platform and disruption lenses. Which explains the case most convincingly, and what evidence would change your view? Keep ownership and corporate scope for Session 9.
Knowledge check
Eleven questions
Answer before you look. Each answer explains itself, and the explanation matters more than the score.
0 / 0 correct
Retrieval practice · 3 minutes
Can you defend the mechanism?
Connect → explain → challenge → revise
Session 8 practice lab
An original diagram and three contexts help you test the mechanism before committing to a recommendation.
A platform grows through useful participation, not counts alone
More useful complements → greater user value → stronger creator incentive → more useful complements. This is a hypothesis to test, not a guaranteed growth loop.
The diagram in words
Useful complements: Solve a real user task.
Value for users: Use, retention and trust.
Incentive for creators: Demand worth serving.
Useful complements improve user value. Retained use can make complement creation worthwhile. Platform rules, access and quality shape every connection; either arrow can fail.
Identify the weakest link, the evidence needed and a condition that would change your recommendation.
Hypothetical · original scenario
Nila Tutor: downloads are the opening, not the outcome
Nila connects lesson creators and working learners. It makes access free for learners, expecting creators to pay for a valuable audience. Learners sign up, but abandon incomplete lessons. The subsidy increased registrations without establishing cross-side value. Nila must first test lesson quality, completed tasks and repeat use; it should also ask whether creators can reach those learners through competing channels.
Discuss: What observation would show that a new creator made existing learners better off, rather than merely increasing the catalogue?
State a claim, explain its causal mechanism and point to evidence. Then identify a credible alternative explanation and one observation that would distinguish it. These prompts invite defensible interpretations; they have no automated score.
Global · NVIDIA / historical context
CUDA: hardware needs work worth running
NVIDIA’s explanation of CUDA describes a programming platform with libraries, tools and a runtime that developers use to accelerate applications. It dates CUDA’s launch to 2006. Source: NVIDIA, “What Is CUDA?” (page dated 10 September 2012; subsequently updated).Teaching inference: useful applications may increase hardware value, and demand may reward developers. That ecosystem mechanism does not alone prove a two-sided marketplace, winner-take-all competition or a disruptive foothold. Use the assigned September 2024 case to test those separate claims.
Discuss: Draw a developer → application → user link. Which link is evidenced in the case, and which still needs testing?
State a claim, explain its causal mechanism and point to evidence. Then identify a credible alternative explanation and one observation that would distinguish it. These prompts invite defensible interpretations; they have no automated score.
Indian · UPI / historical context
UPI: interoperability changes the network boundary
The Indian Economic Service’s UPI explainer dates the pilot launch to 11 April 2016 and describes smartphone transfers between bank accounts and access to multiple accounts through one participating bank’s application. Source: Indian Economic Service, UPI explainer (2016 launch history).Teaching inference: acceptance and usability can reinforce participation, while interoperability changes switching and multi-homing. UPI is a regulated payment infrastructure with multiple participants; do not assume the pricing or monopoly logic of a single private marketplace applies unchanged.
Discuss: Who benefits when more merchants accept interoperable payments? Which participant controls access and trust, and which does not?
State a claim, explain its causal mechanism and point to evidence. Then identify a credible alternative explanation and one observation that would distinguish it. These prompts invite defensible interpretations; they have no automated score.
Four diagnostic questions
Choose an answer, explain it, then use the feedback to revise your reasoning. These are original teaching scenarios and simplified calculations, separate from the existing practice bank.
Practice score: 0 / 4
Run a reversible experiment
Choose one Nila bottleneck. State a prediction, an observable user behavior, a cost boundary and the result that would make you stop. Then compare it with the existing platform studio.
Challenge the NVIDIA interpretation
Use one case page or exhibit to support a platform claim. Ask a partner to offer a sustaining-innovation explanation of the same evidence. Identify what additional observation would discriminate between them.
Use the supplied resources by topic
The folder names reflect an earlier sequence. This map follows the current session topic. Locate these files in your instructor-provided resources or obtain the reading through the licensed course pack, library or publisher; the originals are not rehosted here.
TwoSidedMa.PDF — Eisenmann, Parker & Van Alstyne (2006). Draw the sides and label the benefit behind each arrow; then challenge the subsidy, tipping and envelopment assumptions. Publisher or source record.
TwoSidedMa_co.pdf — Second supplied copy of the platform article. Use as a copy of the same reading, rather than a separate theory or additional assignment. Publisher or source record.
Disruptive Technologies_Catching the Wave.pdf — Bower & Christensen (1995). Identify the customer and performance trajectory that an incumbent’s investment process might overlook. Publisher or source record.
Nvidia, Inc..PDF — Yoffie & von Bargen, NVIDIA case (5 September 2024). Use the historical case evidence to test an ecosystem explanation and a proposed experiment. Keep interpretation distinct from case observation. Publisher or source record.
Preparation
Readings for Session 8
The two assigned readings and the case are copyrighted Harvard Business
School Publishing materials, and the textbooks are in copyright. Get them through the licensed course pack or the library;
use the publisher links below to locate the readings.
Carried forward · Session 7 reading
Kim, W. C., & Mauborgne, R. (2004, October). Blue ocean strategy. Harvard Business Review (Reprint R0410D).
Page numbers on this page follow the 23-page reprint in the course pack. Use this earlier lens to compare value innovation with platform growth and disruption.
Required · the case
Yoffie, D. B., & von Bargen, S. (2024). Nvidia, Inc. in 2024 and the future of AI (HBS No. 9-725-360). Harvard Business School Publishing.
Pages 1–12 and Exhibits 1a, 3 and 7. The same case returns in Session 9 for NVIDIA’s corporate scope.
Further · the tools
Kim, W. C., & Mauborgne, R. (2017). The W. Chan Kim and Renée Mauborgne blue ocean strategy reader. Harvard Business Review Press.
“Value innovation: The strategic logic of high growth” (first published 1997), pp. 1–21, for the value curve and the four questions; “Charting your company’s future” (2002), pp. 93–107, for the strategy canvas.
Further · course textbook
Rothaermel, F. T. (2024). Strategic management (6th ed.). McGraw Hill.
Chapter 7, sections 7.4–7.5, pp. 274–286, on types of innovation and platform strategy; chapter 6, section 6.5, pp. 231–235, on blue oceans and how they go wrong.
Further · course textbook
Grant, R. M. (2019). Contemporary strategy analysis (10th ed.). Wiley.
Chapter 9, “Technology-based industries and the management of innovation”: capturing value from innovation, pp. 222–225, and standards, platforms and network externalities, pp. 232–236.
Carried forward
Porter, M. E. (1985). Competitive advantage: Creating and sustaining superior performance. Free Press.
Chapter 1, p. 20: the passage on pioneering a major innovation, which Session 7 read with the rest of the chapter.
A note on the figures on this page
Every NVIDIA figure comes from the case, with its page or exhibit.
The revenue-mix chart computes segment shares in the browser from Exhibit 1a; “everything else” is revenue less
data centre and gaming. Prices, market shares and performance numbers are the analysts’ and companies’
estimates and claims that the case reports, not audited data. The canvas scores are this page’s reading of the case
and are there to be argued with. Nothing here covers events after the case’s September 2024 date.
Required · Eisenmann, Parker & Van Alstyne (2006)
Strategies for Two-Sided Markets. Harvard Business Review. Prepare a map of both sides, a pricing hypothesis and one reason market tipping might fail.
Reading lens: Publisher record. Applications and model assumptions below are learning exercises.
Required · Bower & Christensen (1995)
Disruptive Technologies: Catching the Wave. Harvard Business Review. Identify whose demands shape resource allocation and why a new market might be overlooked.
Reading lens: Publisher record. Applications and model assumptions below are learning exercises.