ExpertsFMBen Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487]
Invest Like the Best with Patrick O'Shaughnessy · August 18, 2026
5 Selected Lessons:
heuristic · 3 reasoning_pattern · 2
Verification loops determine where AI will actually move the needle first
Source
Invest Like the Best with Patrick O'Shaughnessy — Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487] (August 18, 2026) · at 14:26 ▸ · at 1:01:55 ▸
When this applies
when forecasting which applications of a probabilistic/ML technology will mature fastest
The takeaway
To predict where an emerging capability will have near-term commercial impact, look for existing large-scale verification/feedback markets, not just raw model capability.
Why it works
AI/ML systems improve fastest where there's a fast, cheap, scalable feedback signal for whether an output was good (verifiable domains); domains lacking fast ground-truth signals (long verification loops) will lag regardless of model capability.
One level up
A technology's economic impact is gated as much by the availability of feedback loops to refine it as by the technology's raw power.
Where it stops holding
This stops working when a domain develops a novel way to generate fast verification signals (e.g., via new data sources), collapsing the old verification bottleneck.
Detail
approach: Ask whether a domain has a liquid, scaled feedback mechanism (tests pass/fail, A/B conversion data) before predicting AI will transform it quickly.
signals:
- existence of a large live marketplace generating outcome data (ads, trades)
- bounded/knowable domain like chess/Go vs open-ended domains
- regulatory or access barriers that slow feedback even where data exists (medicine)
rationale: Meta/Google's ad marketplace lets them validate generated content via billions of real human click/purchase decisions — a verification advantage 'no one else' has, explaining why ad-targeting AI may be the biggest near-term AI value driver, more than chatbots.
In their words
They actually can validate their image creation and their text creation in a way no one else can. And their validation is the ad marketplace, running a gazillion A-B tests on all these different things, see what works, see what doesn't.
Hyperscalers threaten NVIDIA via cost-of-capital and fungibility, not raw chip quality
Source
Invest Like the Best with Patrick O'Shaughnessy — Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487] (August 18, 2026) · at 1:09:51 ▸
When this applies
when assessing whether a differentiated-product leader is vulnerable to vertically integrated customers who are also potential competitors
The takeaway
A supplier's differentiation moat erodes fastest when its largest customers both (a) have cheaper capital to build substitutes and (b) can sell those substitutes as commodities without harming their core business.
Why it works
Hyperscalers have structurally lower cost of capital than chip vendors or neoclouds, and because models are increasingly indifferent to the underlying chip, hyperscalers can sell their own chips as commodities without cannibalizing their cloud business (customers aren't drawn to Amazon/Google by chip brand) — undermining the value of NVIDIA's differentiation.
One level up
The biggest long-term threat to a component supplier is usually its own largest customers' vertical integration, enabled by their superior scale and capital access.
Where it stops holding
This stops working when switching from the differentiated supplier imposes real software/ecosystem lock-in costs that remain even as hardware becomes interchangeable (e.g., if CUDA lock-in were stronger).
Detail
approach: Evaluate supplier risk by asking: do my biggest customers have cheaper capital than me, and can they sell competing products as commodities without hurting their core business?
signals:
- customer announces plans to sell internally-developed chips externally (Jassy on Tranium, Google selling TPUs to Anthropic)
- model/software layer becomes chip-agnostic, reducing lock-in (CUDA's moat diminished because models don't care what they run on)
rationale: NVIDIA is selling differentiation; hyperscalers sell their chips as commodities and aren't cannibalizing their cloud attractiveness, which combined with their lower cost of capital makes them NVIDIA's 'biggest problem.'
In their words
They're not selling their chips on differentiation, they're selling their chips as commodities. NVIDIA is the one selling differentiation. People aren't going to Amazon to use tranium. So they're not cannibalizing the attractiveness of their cloud by selling tranium outside.
Free isn't free — marginal inference cost varies by orders of magnitude per user
Source
Invest Like the Best with Patrick O'Shaughnessy — Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487] (August 18, 2026) · at 6:01 ▸ · at 20:23 ▸ · at 25:26 ▸
When this applies
when a product is marketed as 'free' but has highly variable serving costs across users
The takeaway
Before accepting a 'zero marginal cost' or 'free' framing, check whether usage intensity varies enough across customers to break that assumption.
Why it works
Inference/serving cost scales with how much compute a given user session consumes (e.g., test-time reasoning length), so 'average cost' hides a huge spread between light and heavy users.
One level up
Aggregate unit-economics claims can mask heterogeneity that determines whether a business model actually works.
Where it stops holding
This stops working when usage is genuinely uniform across users (e.g., simple static content) so marginal cost truly is near zero for everyone.
Detail
rule: Don't treat AI inference cost as one number; segment by usage pattern (casual chat vs. long test-time-scaling reasoning) before judging margins or 'free' claims.
threshold: n/a - qualitative spread, not a single number
rationale: The marginal cost question for a light user and a heavy test-time-scaling user 'are not even remotely in the same universe.'
In their words
The marginal cost question for the different user, the user using free chat GPT and the user trying to solve a math theorem, they're not even remotely in the same universe.
Advertising monetizes better than subscriptions because advertisers absorb price increases
Source
Invest Like the Best with Patrick O'Shaughnessy — Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487] (August 18, 2026) · at 25:26 ▸
When this applies
when choosing between charging end users directly versus a third-party-funded (ad) model for a consumer product
The takeaway
Where consumer willingness-to-pay is capped, monetize through a third party whose demand for your reach scales independently of consumer price sensitivity.
Why it works
Raising subscription prices directly triggers consumer price-elasticity and churn risk; raising ad prices shifts the burden to advertisers competing for the same attention, with far less consumer-facing friction.
One level up
The entity bearing the price increase determines how much pricing power a business truly has — choose your payer to maximize that elasticity slack.
Where it stops holding
This stops working when the product lacks sufficient scale/attention to build a real ad marketplace, or when the audience is too small/niche to attract advertiser demand.
Detail
rule: For consumer products, prefer advertising over subscription pricing because the advertiser, not the consumer, bears cost increases.
threshold: n/a
rationale: Charging people money is hard; giving people things for free (ad-supported) is easy, and ad monetization scales with reach rather than consumer tolerance for price hikes.
In their words
The thing about advertising with consumers is your ability to monetize the consumer goes up as the bottom because the advertiser is bearing the price increase. So there's zero elasticity issues.
Companies should double down on what they're structurally good at, not chase every new paradigm
Source
Invest Like the Best with Patrick O'Shaughnessy — Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487] (August 18, 2026) · at 47:48 ▸
When this applies
when deciding whether to chase a hot new paradigm that conflicts with your organization's core operating DNA
The takeaway
Before assuming a company 'must' pursue a new paradigm, test whether the paradigm's nature matches the organization's demonstrated operating strengths, not just its market position.
Why it works
A company's culture, processes, and decision-making habits (e.g., deterministic, zero-defect-tolerant supply chain discipline) are optimized for one kind of problem and are often poorly suited to a fundamentally different kind (probabilistic, iterative AI development).
One level up
Comparative advantage, not FOMO, should determine which battles an organization fights.
Where it stops holding
This stops working when the new paradigm becomes so central to the core product that non-participation is existential rather than a choice (e.g., if AI quality becomes the deciding factor in phone purchases).
Detail
rule: Judge whether a company should pursue a new technology by asking 'on what basis should we expect them to be good at this,' given their demonstrated operating model.
threshold: qualitative: deterministic vs probabilistic mismatch
rationale: Apple ships physical products once and can't tolerate failure (no iPhone recalls ever); AI is inherently probabilistic and iterative — a different discipline entirely.
In their words
At the most crude level, AI is this probabilistic endeavor. Apple is the king of deterministic products.