ExpertsFMAnthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers
Sharp Tech with Ben Thompson · October 30, 2026
5 Selected Lessons:
heuristic · 2 case · 1 framework · 1 priority · 1
Live in the present rather than obsessing over an uncontrollable future outcome
Source
Sharp Tech with Ben Thompson — Anthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers (October 30, 2026) · at 1:31:40 ▸ · at 1:35:45 ▸
When this applies
when facing a long, high-stakes, largely uncontrollable process (career bet, long project, societal shift) with an uncertain outcome
The takeaway
When you have no control over an outcome, maximize value from the present process itself rather than deferring well-being to a contingent future result.
Why it works
Fixating on an uncontrollable future outcome produces chronic anxiety without changing the outcome, whereas treating the present process as the source of value lets a person extract enjoyment regardless of the eventual result and insulates them from outcome-dependent misery.
One level up
Separate what you control (your present engagement, effort, enjoyment) from what you don't (eventual outcomes), and anchor your well-being to the former.
Where it stops holding
This stops working when the present process itself is genuinely miserable or when some degree of outcome-anxiety is functionally necessary to drive the actions needed to influence the outcome.
Detail
hierarchy:
- Control what you control: your own engagement, presence, and enjoyment
- Accept you don't control outcomes (sports results, AI's future, markets)
- Extract value from the process/journey regardless of eventual outcome
- Deal with the future when it arrives rather than living in dread of it now
rationale: The worst thing about being emotionally invested in sports is you have no control; the same logic applies to worrying about an unknown AI future — better to live in the present, be nice to people, learn and explore, and figure out the future when it gets there.
In their words
How about you enjoy life right now? How about you live your life instead of always thinking in the future or thinking in the past? Live in the present.
Craig Moffett's 'is the money there' CapEx test
Source
Sharp Tech with Ben Thompson — Anthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers (October 30, 2026) · at 42:15 ▸
When this applies
when evaluating any capital-intensive buildout cycle (infrastructure, platforms, networks) for sustainability
The takeaway
To judge sustainability of a capital buildout, sum total external revenue entering the system and compare it to total capital cost of the whole value chain, ignoring internal layer-to-layer transfers as noise.
Why it works
Intermediate revenue flows between layers of a value chain are internal transfers and don't address whether the system as a whole generates enough external revenue to cover the total capital deployed; only end-demand inflows matter for solvency.
One level up
When analyzing any multi-layer value chain's health, abstract away internal transactions and look only at net inflows versus total system cost.
Where it stops holding
This stops working when internal transfer pricing materially distorts incentives (e.g. one layer subsidizing another to inflate apparent demand), in which case gross inflow figures may themselves be unreliable.
Detail
procedure:
- Identify total CapEx committed across the entire value chain (chips, clouds, labs)
- Identify total external revenue entering the system from end customers (consumers/enterprises)
- Ignore revenue transfers between layers of the chain as internal noise
- Compare total external inflow to total system cost over the relevant time horizon
benchmarks: {"current_state":"Anthropic's revenues are impressive but nowhere close to covering total CapEx being deployed across the ecosystem"}
rationale: All that matters is whether enough money is coming into the system to support the system; everything else is distributional noise.
In their words
All that really matters is enough money coming into the system to support the system.
R&D cost growing faster than serving cost signals a structural, not temporary, cost problem
Source
Sharp Tech with Ben Thompson — Anthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers (October 30, 2026) · at 6:47 ▸
When this applies
when assessing any business whose core input cost (R&D, content, inventory) is rising faster than its unit cost-to-serve, driven by competitive necessity rather than scale
The takeaway
A cost that scales with competitive necessity rather than with your own operational choices is not something you can 'solve' — it is an ongoing structural feature of the business.
Why it works
When a cost category is driven by competitive racing dynamics (must get better to stay relevant) rather than operational scale, the company loses control over the pace and size of that spend — it becomes exogenous rather than a fixable, one-time problem.
One level up
Always ask whether a rising cost is inside or outside your control before judging whether margins will improve over time.
Where it stops holding
This stops working when the competitive dynamic cools (e.g., the frontier race ends or a company stops needing to chase it), at which point the cost becomes manageable.
Detail
rule: If R&D/training spend grows faster than cost-to-serve even while usage explodes, the spend is driven by competition, not by serving customers, and is structurally uncontrollable.
threshold: Training cost > inference cost as the largest cost category, growing faster than usage-driven serving costs.
rationale: They need a better model because of competition pacing the frontier; this spend is not discretionary.
In their words
The largest portion of their cost is not in fact inference. The largest portion of their cost is this training cost, and the speed with which they need to spend money on that is out of their control.
Multiplying humans (top-line growth) beats replacing humans (bottom-line cost cuts) as an AI adoption strategy, but takes longer and favors new entrants
Source
Sharp Tech with Ben Thompson — Anthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers (October 30, 2026) · at 48:54 ▸
When this applies
when deciding whether to deploy a new capability to cut costs internally or to create new offerings/workflows
The takeaway
Transformative technologies more reliably create new value by enabling new entrants to build fit-for-purpose workflows than by incumbents retrofitting it for cost-cutting — but this path is slower and timing risk is the main constraint.
Why it works
Cost-cutting applications of a new technology are usually bolted onto existing structures (SaaS incumbents) that weren't built for it, limiting gains, whereas revenue-multiplying applications require building new workflows from scratch, which new AI-native entrants can do better but takes years to displace incumbents.
One level up
When judging how a general-purpose technology will diffuse, weigh 'multiplying' use cases built by new entrants more heavily than 'replacing' use cases bolted onto incumbents, but discount for the time such diffusion takes.
Where it stops holding
This stops working when incumbents can restructure fast enough to capture multiplying use cases themselves, or when the financing for new entrants runs out before diffusion completes.
Detail
rule: Favor 'multiplying humans' (new workflow creation) over 'replacing humans' (cost cutting) as the more durable AI value-creation path, but expect it to take years and come from new companies rather than incumbents.
threshold: n/a - qualitative pattern, but explicitly tied to 'takes longer' and 'comes from new companies' as the two defining frictions
rationale: SaaS incumbents are not fundamentally constructed to multiply humans; AI-native companies replacing them is how the transition happens, but this is slow while CapEx is being spent now.
In their words
There's this this this difference between multiplying versus replacing. And when you're multiplying humans, you're basically creating entirely new things to do, new workflows, new opportunities.
Anthropic's heterogeneous infrastructure bet pays off
Source
Sharp Tech with Ben Thompson — Anthropic Opens the Kimono, Bubble Questions and Frontier Costs, Q&A on Meta, Starlink, Brewers (October 30, 2026) · at 9:58 ▸
When this applies
when a company is forced by a key investor/partner into using a non-preferred technology stack early on
The takeaway
A constraint imposed early (multi-platform support) can become a long-term structural advantage via the abstraction layer it forces you to build.
Why it works
Being forced to support multiple backends early creates an abstraction layer that becomes a durable flexibility advantage, letting the company add further chips/clouds cheaply and avoid lock-in to any single platform.
One level up
Forced diversification, if survived, often outperforms chosen specialization because it builds optionality you didn't know you'd need.
Where it stops holding
This stops working when supporting multiple platforms is prohibitively costly relative to the company's scale, making the abstraction layer a drag rather than an asset.
Detail
situation: Amazon invested in Anthropic and required them to use Trainium, a chip Anthropic didn't want.
decision: Anthropic built infrastructure to be heterogeneous across chips and clouds rather than resisting.
outcome: This flexibility let Anthropic win enterprise customers on AWS/Google/Microsoft that OpenAI, locked to Microsoft, couldn't as easily reach.
lesson: An externally imposed technical constraint, if absorbed into architecture, can become a competitive moat.
In their words
They've built a layer of abstraction that is a very good place to be. I think it's benefited very well.