MBUS 853 — Session 3: Digital Landscape

Block 1 — Session Theme: The Landscape Just Got Faster

From Leadership Execution to Landscape Speed

Session 2 asked what leadership behaviors make or break a digital transformation. Session 3 asks a question underneath that one: even with perfect execution, how long does a digital or AI advantage actually last before competitors catch up? DeepSeek is the sharpest possible test case — a Chinese quant-fund spinoff that, by its own account, matched or beat Silicon Valley's leading models at a small fraction of the reported training cost, open-sourced it, and erased $589 billion of Nvidia's market capitalization in a single trading day.

A capability that took a rival 15 years and hundreds of millions of dollars to build can now be matched, in some dimensions, in months and for a fraction of the cost.

This session tests Session 1's DBS optimism against a landscape where AI capability itself may be the thing that commoditizes fastest.

Why This Case Follows GE, Not Just DBS

DeepSeek isn't just a landscape story — it's also a leadership story in miniature. Founder Liang Wenfeng's stated ambition ("we're done following... it's time to lead") echoes Immelt's ambition to make GE a "top ten software company." The difference worth interrogating in discussion: DeepSeek moved from insight to shipped, adopted product in roughly two years, operating with startup speed and a tightly integrated team, versus GE's multi-year, multi-CEO platform build. Landscape speed rewards structural agility as much as it rewards capital.

Block 2 — Article: AI Won't Give You a New Sustainable Advantage (Barney & Reeves, HBR 2024)

The Core Claim: Value Creation Is Not Value Capture

Historically, transformative general-purpose technologies (the steam engine, the electric motor, the personal computer) rarely became a sustained source of advantage for any one adopter — precisely because their impact was so large that virtually every competitor was compelled to adopt them too, often erasing incumbents' prior advantages instead. Barney and Reeves argue generative AI fits this pattern exactly, for a reason unique to how the technology learns.

Claim 1

First-mover advantage erodes itself

Because gen AI is trained on constantly updated data, your first-mover use of it becomes visible (through your public actions or results) to the same models your competitors query next. Late movers benefit from your prior effort, not just their own.

Claim 2

You probably can't out-build the platform

Building a custom, general-purpose gen-AI platform to rival OpenAI or Google is rarely rational — those firms have years of scaling experience, and any custom system a company does build will likely be matched by a competitor's own build, partnership, or purchase.

Claim 3

Proprietary data is a weaker moat than it looks

Competitors often hold functionally equivalent (if not identical) datasets. Beyond a certain sample size, more data stops improving pattern recognition. And proprietary datasets are notoriously leaky — "you could be one disgruntled employee away from having yours shared with the entire world."

Claim 4 — the silver lining

AI amplifies advantages you already have

If you already hold rare, hard-to-imitate resources or capabilities (Amazon's supplier network, warehousing, and logistics stack is the article's example), applying gen AI to those specific assets can compound an existing moat — because a rival without the underlying asset can't replicate the AI-amplified result either.

"The only ones that will actually win with it will be those that can apply it to amplify the advantages they already have."

— Barney & Reeves — the article's closing line, and the sharpest lens to apply to DeepSeek itself

The DeepSeek twist: DeepSeek's own founding story is almost a direct test of Claim 4. Liang Wenfeng didn't start as an AI researcher — he ran a quantitative hedge fund (High-Flyer, over 100 billion yuan under management) and had already built infrastructure for large-scale GPU compute and algorithmic pattern recognition for trading. DeepSeek arguably amplified a pre-existing, hard-to-imitate capability (quant infrastructure + compute-efficient engineering culture) rather than starting from zero.

Block 3 — Case: DeepSeek: The Emergence and Evolution of an AI Technology

A $589 Billion Trading Day

DeepSeek, launched January 2025 by Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd., introduced an open-source LLM that markets read as a credible substitute for Silicon Valley's leading closed models — at what the company implied was a fraction of the training cost (under $6 million in Nvidia chip compute, versus $100M+ reportedly required for OpenAI's latest offering at the time). Within a week of its US launch, DeepSeek's app had 3 million downloads and was the #1 free app on Apple's iPhone store. The market reaction was immediate: the S&P 500 fell 1.5%, the Nasdaq tech index fell over 3%, and Nvidia alone lost 16.9% of its value — $589 billion, the largest single-company market-cap loss in history at that point.

From Hedge Fund to AI Lab

Liang Wenfeng, a Zhejiang University computer science graduate, co-founded the High-Flyer quantitative hedge fund in 2015 (over 100 billion yuan / $13.79B AUM by 2021), using machine learning to refine trading strategies. While running High-Flyer, he began buying Nvidia chips for an AI side project, eventually reporting a cluster of 10,000 Nvidia A100 GPUs as DeepSeek's compute foundation. DeepSeek operated under High-Flyer with an explicit goal of advancing artificial general intelligence, and deliberately hired liberal arts graduates alongside engineers to curate training data and improve the linguistic and cultural nuance of its Chinese-language outputs — a differentiated talent strategy aimed at reducing reliance on Western infrastructure and data.

The AI Value Chain and Where DeepSeek Sits

Stage	What Happens	Who Dominates
Data Collection & Management	Gathering, storing, and managing training data at scale	Chipmakers (Nvidia, AMD) and cloud platforms (AWS, Azure, Google Cloud) — infrastructure gatekeepers
Model Development	Training transformer-based LLMs to understand and generate language	OpenAI, Google, Meta, Anthropic, and now DeepSeek — a widening but still concentrated field
Model Deployment	Productizing models into apps, APIs, copilots, and vertical tools	The most dynamic, commercially diverse layer — low barriers via APIs and fine-tuning let startups compete here

DeepSeek's disruption is specifically a Model Development-layer disruption: it challenged the assumption (built into how OpenAI, Google, and Meta had all been racing) that model quality is a direct function of training cost and compute intensity. By substituting engineering cleverness for scale, DeepSeek called that entire cost structure into question — which is precisely why chip demand, not just model competition, was what markets repriced first.

Competitor Reactions — A Spectrum of Concern

Openly Rattled

Investors: Nvidia -16.9% in a day. Cohere's founders publicly began exploring how to prove "the key to AI development is innovation and efficiency, not excessive compute" — an implicit concession that DeepSeek had reframed the competitive question.

Competitively Engaged

OpenAI's Sam Altman called R1 "impressive... for the price" while insisting "more compute is more important now than ever." Meta assembled four engineering teams specifically to analyze DeepSeek's methods for Llama.

Downplaying

Google DeepMind's Demis Hassabis called it strong engineering that used known techniques rather than a new scientific breakthrough. Anthropic's response was similar — it compared DeepSeek's performance to US models roughly seven to ten months older, framing the gap as expected progress rather than disruption (the case's own paraphrase of both companies' positions, not direct quotes from either).

Reframing Entirely

Meta's Yann LeCun: "To people who see the performance of DeepSeek and think 'China is surpassing the US in AI,' you are reading this wrong. The correct reading is: open-source models are surpassing proprietary ones."

The Open Question DeepSeek Still Faces

OpenAI accused DeepSeek of using knowledge distillation from OpenAI's own models to improve its training — a technique the case notes is common industry practice, not unique misconduct. Geopolitically, DeepSeek was banned from some government devices in the US and faced data-privacy scrutiny in the EU and Italy over how user data was handled, and members of the US Congress moved to introduce a bill banning it from government devices entirely. The case's own closing tension: having proven the disruption is possible, can DeepSeek convert a viral, low-cost technical breakthrough into a scalable, sustainable, widely-adopted business — or does its own playbook (cheap, fast, open-source replication) get turned against it by the next fast-follower?

Block 4 — Memo Protocol: First Team Case Study Memo Is Due This Week

Format Reminder Before the Team Writes

Per Appendix A of the syllabus, this memo is due Thursday 11:59pm before Session 3, based only on the DeepSeek case (no article or outside sources), submitted as a two-page Word document, 11-point font, written wholly by the team in the team's own words. This block is a structure check, not a draft — the actual issues, alternatives, criteria, and recommendation need to be worked out and written by the team itself.

To Name the one main decision maker in the case — for DeepSeek that's almost certainly Liang Wenfeng, not a committee or "DeepSeek" as an entity.

Issues Exactly 5, each a 3–5 word sub-heading + 20–30 word description, each tied to a specific case fact or figure (not a general statement).

Problem/Decision 40–60 words: the one or two underlying causes behind most of the 5 issues, ending in a decision framed as a question.

Alternatives Exactly 3, mutually exclusive, feasible, going-forward — not simultaneous initiatives, and not status quo.

Criteria Exactly 3 standards (financial / strategic / executional, etc.) used to judge the alternatives against each other.

Evaluation/Recommendation 120–140 words, pros/cons of each alternative against each criterion, no table, ending in a justified pick.

Actions Exactly 3, each a step the company hasn't already taken in the case, needed to execute the recommendation.

Case-only constraint, worth double-checking before submission: the DeepSeek case is unusually dense with named competitor reactions (Altman, Hassabis, Zuckerberg, LeCun) — it's easy to accidentally reach for outside knowledge about what those companies did after the case's own timeline ends. Stay inside what the case states as of its own close.

Academic integrity — GenAI is banned in submitted work for this course. This prep page (and the diagnostic/discussion analysis below) is study material for class discussion, not memo text — it must not be copied into the team's submission. The memo itself has to be written, sentence by sentence, by the team.

Block 5 — Case Diagnostic: Issues, Decision, Position (Discussion Prep)

Applying the Case Prep Protocol

This is discussion prep for class, structured the same way Sessions 1–2 were — it is deliberately lighter than a full memo (no alternatives/criteria/actions triad) so it stays a discussion aid rather than a draft of the team's actual submission.

Step 1 — Who and What

Decision maker: Liang Wenfeng, founder of DeepSeek (operating under parent company High-Flyer). Core challenge: DeepSeek proved it could produce a near-frontier model at a fraction of the presumed cost and distribute it for free — but per Barney & Reeves' framework, an advantage this visible and this easy to replicate (open-source, published techniques, cheap to reproduce) is structurally the hardest kind to sustain.

Step 2 — Candidate Issues Grounded in Case Facts

The moat is publicly legible. DeepSeek's cost and technique advantages were disclosed and widely reported — precisely the "value created but not captured" problem the article describes, since competitors can now target the same efficiency gains.
Open-source undercuts DeepSeek's own monetization. 3 million downloads in under a week proves demand, but an open architecture that anyone can run locally limits the direct revenue DeepSeek itself can capture from that demand.
Geopolitical exposure is mounting. Government-device bans, a proposed US congressional bill, and EU/Italian data-privacy scrutiny all threaten DeepSeek's addressable market independent of its technical performance.
Knowledge distillation accusations threaten legitimacy. OpenAI's claim that DeepSeek used distillation from its models — common industry practice or not — creates a legal and reputational overhang DeepSeek has to manage while scaling.
The efficiency narrative is a two-edged sword. If DeepSeek's core claim (frontier performance without frontier compute) is true, it also lowers the barrier for the next fast-follower to replicate DeepSeek's own approach.

Step 3 — A Position

Underlying problem, one sentence: DeepSeek won attention and adoption by making its advantage maximally visible and reproducible (open-source, published efficiency claims) — the same choice that makes the advantage nearly impossible to sustain under Barney & Reeves' framework, since gen AI advantages erode fastest when they are public and easy to copy.

Counterargument to weigh: One could argue DeepSeek doesn't need a sustained model-layer advantage at all — per the article's "silver lining," the durable asset might be High-Flyer's underlying quant-trading infrastructure and compute-efficiency culture, which a competitor can't replicate just by copying DeepSeek's published papers. The strongest response has to weigh whether that underlying capability is rare enough, and hard enough to imitate, to count as the kind of asset the article says AI can actually amplify — versus whether it's ultimately just engineering talent that could walk out the door.

Second counterargument — the real threat isn't economic: Barney & Reeves' whole framework assumes competitors erode your advantage by copying it. DeepSeek's case shows a second, non-economic erosion path the article doesn't address: government bans on state devices, a proposed congressional bill, and EU/Italian data-privacy probes can cut off market access regardless of how durable or rare DeepSeek's underlying capability is. On this reading, debating whether High-Flyer's infrastructure is a defensible moat may be answering the wrong question — the more urgent constraint on DeepSeek's "way forward" is geopolitical, not competitive-strategy, and no amount of asset rarity fixes a government device ban.

Step 4 — 30-Second Cold-Call Answer

DeepSeek proves Barney and Reeves right in real time, just not in the direction most people assume: it didn't sustain an advantage over OpenAI or Google, it triggered instant imitation — Meta stood up four engineering teams, Cohere's founders publicly pivoted to an efficiency narrative, all within days of the January 27 launch. That's because DeepSeek made its advantage maximally visible and reproducible by open-sourcing it, which is exactly the condition under which the article says gen-AI advantages evaporate fastest. The one asset in this case that actually fits the article's "silver lining" is High-Flyer's quant-trading infrastructure — years of GPU compute experience and an efficiency-first engineering culture that a competitor can't get just by reading DeepSeek's papers. So the real strategic question isn't "is DeepSeek's model good" — it already proved that — it's whether Liang Wenfeng builds the business around that upstream capability before geopolitics (the government-device bans, the congressional bill) closes off the market before the economics even get tested.

Block 6 — Discussion Questions & Sharp Answers

Likely Professor Questions

Framing to expect: (1) How did DeepSeek disrupt the AI landscape? (2) Is DeepSeek's advantage sustainable? (3) What should incumbent AI labs do in response?

Q1: Barney & Reeves say gen AI rarely sustains an advantage. Does DeepSeek prove or disprove that thesis?

It largely proves it, on the article's own terms — but in an unusual direction. DeepSeek didn't sustain an advantage over OpenAI/Google/Meta; it briefly held one and immediately triggered a wave of imitation (Meta's four engineering teams, Cohere's public efficiency pivot). What DeepSeek actually demonstrates is the article's core mechanic in real time: a visible technical achievement gets absorbed into the competitive dataset almost instantly, exactly as the "toothbrush" thought experiment predicts.

Owo's data/analytics approach for NGX retail investors works today partly because it's a small, under-served market few larger players have bothered to build for — a reminder that visibility is itself a strategic choice, not just a byproduct of success.

Q2: Was open-sourcing DeepSeek's model a strategic mistake, given it gave away the advantage almost immediately?

Not necessarily a mistake — it may have been the correct read of a landscape where the advantage couldn't have stayed proprietary for long anyway. If the underlying insight (efficiency over raw compute) was always going to leak or be independently discovered, open-sourcing it first converts an unsustainable technical lead into a sustainable reputational and ecosystem lead — DeepSeek becomes the name associated with the shift, the way Linux's early open-sourcing built durable community and enterprise trust even as the code itself was freely copyable.

Stutern's B2B SaaS platform succeeded partly by being generous with skills-matching data and insights to partner businesses — building ecosystem trust mattered more than guarding every technical detail.

Q3: Given the article's "silver lining," what pre-existing asset should DeepSeek actually be trying to amplify with AI, rather than competing head-on with OpenAI's model-quality race?

High-Flyer's quantitative trading infrastructure — years of large-scale GPU compute experience, algorithmic pattern-recognition discipline, and a culture built around extracting more signal per dollar of compute than rivals. That's a genuinely rare, hard-to-imitate combination (a hedge fund with AI-lab-grade infrastructure is not a common starting point), and it's the one asset in the case that a fast-follower copying DeepSeek's published papers cannot simply replicate.

This is the strongest reframe to lead class discussion with — it shifts the conversation from "is DeepSeek's model advantage sustainable" (no, per the article) to "what actually is DeepSeek's rare asset" (yes, arguably) — which is a more analytically interesting question and directly applies the article's own framework rather than just its headline.

Block 7 — Participation Hooks & Taju's Edge

How to Contribute Distinctively

Team Class Contribution (20%) rewards quality and diversity of voice, not volume — and this is the first session running alongside a graded memo, so participation and memo prep should reinforce each other rather than duplicate the exact same points verbatim.

Open Strong

Don't open with "DeepSeek disrupted the AI market." Open with the mechanism: DeepSeek didn't just build a cheaper model — it made the cost structure of frontier AI visible and public, which is what actually moved $589B out of Nvidia in a day.

Push the Consensus

Class will likely say "DeepSeek proves China can compete in AI." Push further, using LeCun's own framing from the case: the more precise claim is that open-source is what's actually surpassing proprietary models — nationality is a distraction from the real structural story.

Bridge to Session 2

DeepSeek moved from insight to global adoption in roughly two years with a lean, integrated team — GE spent nearly a decade and 5,500 hires to get GE Digital to a fraction of that market impact. Landscape speed rewards organizational agility as much as capital.

Taju's Edge — Owo

Building a stock-valuation tool for NGX retail investors is a live bet that a compute-efficient, narrowly-scoped analytics product can compete with well-resourced incumbents by targeting an underserved niche — DeepSeek's playbook at a much smaller scale.

Taju's Edge — Redamo Labs

Analytics frameworks built for enterprise IAM improved verification accuracy by working within existing infrastructure constraints rather than assuming unlimited compute or data — the efficiency-over-scale discipline DeepSeek is built on.

Taju's Edge — Prodigy Education

A/B testing and predictive models across a 3M+ user platform showed that targeted, well-designed experimentation can outperform brute-force scale — directly relevant to the article's "bigger dataset isn't necessarily better" claim.

Block 8 — Reflections Journal Prep (Fill In After Class)

Concept Must Come From Live Slides

As with Session 2, this prep only has the article and case — not Prof. Mufti's Session 3 slide deck — so the journal's concept half has to be captured live in class, not guessed at here. The example half (from professional experience, never from the article or case) is ready to pair with whatever concept the session actually covers.

Concept — Fill In From Session 3 Slides

[3–7 word identification — write after class]

150–200 words, from class slides only. Landscape-focused candidates to listen for: competitive advantage erosion, technology diffusion curves, first-mover vs. fast-follower dynamics, or moat/defensibility frameworks — write down whichever the professor actually presents.

Candidate Example — Ready to Pair With Whatever Concept Fits

Owo — Choosing a Defensible Niche Over a Crowded Race

Building a stock-valuation tool for Nigerian Stock Exchange retail investors meant deliberately not competing head-on with well-resourced global fintech products covering US and European markets. The insight wasn't proprietary technology — most of the underlying valuation techniques are well known — it was recognizing that NGX-specific data, local market context, and retail-investor trust in that specific market were the actual scarce assets worth building around. A larger competitor could copy the interface in weeks; replicating the local market relationships and data curation would take much longer. That distinction — between an easily copyable technique and a genuinely scarce underlying asset — is the difference between a temporary edge and a durable one, and it only became clear in hindsight, after choosing the niche rather than the broader market.

Block 9 — Key Takeaways

What to Walk Away Knowing

Visible advantages erode fastest. DeepSeek's public, open-source disclosure of its efficiency techniques is the clearest possible illustration of "value created but not captured."

Scale of data or compute is not automatically a moat. Beyond a certain threshold, more data or bigger models don't reliably translate into a harder-to-copy advantage — DeepSeek's entire premise was proving this empirically.

The durable asset is usually upstream of the AI application itself. High-Flyer's quant infrastructure, not DeepSeek's model weights, is the harder-to-replicate resource — a direct application of the article's "silver lining."

Landscape speed now rewards organizational agility as much as capital. DeepSeek reached global relevance faster and leaner than GE Digital did with orders of magnitude more resources — a direct callback to Session 2's execution themes.

Looking Ahead — Session 4: Digital Organization

→ Session 4 (Organization)

Session 3 asks "how fast does an AI advantage erode?" Session 4 (Building the AI-Powered Organization; Pernod Ricard case) asks "what organizational structure and culture let a company keep generating new advantages faster than they erode?"

↔ Recurring Thread: Amplify, Don't Chase

Barney & Reeves' "amplify your existing advantage" principle recurs through Sessions 4–5 (Organization, Architecture) as the underlying design question: what should a digital/AI investment be built on top of, rather than built from scratch.

Memo #2 Due Before Session 4

The team's second Case Study Memo (Pernod Ricard) is due the Thursday before Session 4 at 11:59pm — confirm the case source material is available with enough runway to draft as a team, not the night before.