The Landlord in the Loop

Marta built a small content studio. Six months of work on prompts, templates, and voice conditioning. Three paying clients, fifteen hundred dollars a month in steady revenue, ChatGPT Plus and Midjourney and Canva as the tool stack, a margin structure that finally let her quit the agency job she had been saving to leave. Then OpenAI changed the rate limits on her plan. The thing she had been doing — fast iteration against long-context drafts for the same three clients — stopped working the way it had worked the week before. Her workflows slowed. Her margins compressed. Her productivity advantage, the thing she had actually been selling, lived on an infrastructure she did not own and could not negotiate with.

Marta is a composite — stitched together from the dozens of solopreneur stories that friends and former students have relayed over the past twelve months. When people describe what happened to her, they say she lost her business to AI. That description gets part of it right. AI did substitute for the labor her clients would have paid her human-sized premium for in 2022. But the description also misses something material. What happened more precisely is that the savings from the substitution stopped flowing to her. The platform repriced the capability, and the capability flowed through the platform rather than through her. The savings have been redirected, not simply destroyed.

The current discourse on AI and labor has bundled two mechanisms that need to be pulled apart. The first is capability displacement: the fact that AI systems now perform at near-zero marginal cost tasks that used to require paid human labor. The second is platform-mediated value capture: the structural feature of the AI economy in which those capabilities flow through a small number of centralized platforms that set pricing, enforce policies, and appropriate most of the economic surplus capability substitution creates. Both mechanisms are real. They compound. They are separate. The conflation has obscured where the actual pain is being inflicted, on whom, and — most importantly for entrepreneurs — what alternative configurations are available.

I want to argue that the conflation is not incidental. It lets platforms offload the political cost of extraction onto a technology narrative, and it lets workers-turned-entrepreneurs mistake a distributional problem for a capability problem. Disentangling the two changes what the “AI displacement problem” is and therefore what can be done about it. If the problem is capability displacement, the question is how workers adapt to an economy in which many tasks automate away. That framing produces familiar answers: retraining, education, UBI, social insurance. If the problem is platform-mediated value capture, the question is how the economic surplus from AI-driven productivity gains distributes among platforms, workers-turned-entrepreneurs, end users, and the public. That is a political-economy question with a different answer set: antitrust, open-weights policy, infrastructure investment, strategic stack choices by entrepreneurs. Neither framing alone is adequate. Both together are what this moment requires.

Two Mechanisms

Capability displacement is the substitution effect. When a language model summarizes a document for a hundredth of a cent, the firm that used to pay a junior analyst fifty dollars an hour for the same summary pays less. The task reclassifies from human labor to computed output. This pattern is real, it is accelerating, and the capability envelope keeps expanding. Stanford's 2026 AI Index shows the top frontier labs now cluster within a narrow band on Arena Elo and adjacent reasoning benchmarks — a far tighter cluster than eighteen months earlier.Stanford AI Index^[1] What was expert-level knowledge work in 2024 is now at the margin of what a mid-tier model does on demand. No amount of platform reform will reverse this.

Platform-mediated value capture is the distributional effect. When the firm that used to pay Marta now pays a platform for the summarization capability, the wages she was collecting do not evaporate into consumer surplus. They migrate. They become platform revenue. Call it the landlord in the loop: the platform sits silently inside every transaction the capability enables, collecting a share of every flow, and the share is set by rules the entrepreneur cannot contest. The platform's share of those savings is negotiated from a position of considerable structural advantage — the capability flows only through the platform, switching costs are real, and the platform holds the pricing discretion of a near-monopsony over the specific capability class. Anthropic, OpenAI, Google, and a small set of others currently set the terms on which capability displacement translates into end-user cost reduction. PwC's 2026 AI performance study found that approximately three-quarters of the economic gains from AI are captured by roughly the top twenty percent of firms.PwC^[2] That concentration reflects platform architecture more than capability differentials — the capability gap between the leading lab and the fourth lab has compressed to a few percentage points. The concentration is produced by the architecture, not the capability.

The analytical wedge between the two mechanisms is a counterfactual. Hold capability constant. Vary the platform architecture. What changes? Under centralized platforms, an entrepreneur who builds on summarization, content generation, or analysis pays the platform per query and takes a margin on the value she delivers to end customers. Her economics are bounded by the platform's pricing and policy. Under open-weight architectures, the same entrepreneur runs the same capability on her own infrastructure. She pays fixed compute costs rather than per-query costs. The platform cannot raise her marginal cost by changing its pricing. The platform cannot deprecate her workflow by updating its model. The platform cannot decide that her use case now violates its terms of service. The capability-displacement pressure is identical. The platform-dependence pressure is zero.

This is no longer a hypothetical counterfactual. In the last twelve months, the global AI market has quietly executed one of the fastest architectural shifts I can recall in years of studying technology adoption. The shift has been happening beneath the surface of the public discourse, and naming it is the first step toward understanding what it means for entrepreneurs.

The Empirical Wedge

The current moment is the first in which the capability-versus-platform distinction can be tested empirically rather than inferred historically. Three developments make the test possible.

1. The capability plateau at the frontier

The four major Western labs — Anthropic, OpenAI, Google DeepMind, and xAI — now cluster within a narrow band on the reasoning benchmarks most relevant to labor-displacement tasks.^[1] Whatever differentiation used to live at the capability layer has substantially collapsed. The competition has moved to the distribution layer, the enterprise layer, the integration layer. The models themselves are converging.

2. The rise of near-capability-equivalent open-weight models

This is the part of the story that deserves more attention than it has received. Chinese open-weight models have gone from roughly 1.2 percent of global AI usage in late 2024 to nearly thirty percent a year later — one of the fastest shifts in deployment architecture I have seen in the AI era.OpenRouter^[3] Alibaba's Qwen family replaced Meta's Llama as the most-downloaded language model family on Hugging Face by late 2025, with tens of thousands of derivative models spawning across the Qwen line alone.^[4]

DeepSeek V3 has pushed capability parity with Western frontier labs on several technical benchmarks at a small fraction of the deployment cost; the forthcoming V4 release, still unlaunched at time of writing, is expected to push further. Moonshot's Kimi K2 and Alibaba's Qwen3-Max price input tokens at a fraction of what the major Western closed labs charge, with the exact spread depending on whether cached or standard rates are applied at the provider level. Andreessen Horowitz partner Martin Casado has reported that among early-2026 startups pitching with open-source AI stacks, roughly eighty percent of those stacks are running on Chinese models — an eighty-percent share of the open-source subset, not of all startup stacks.^[5] Airbnb's Brian Chesky has publicly stated a preference for Qwen because it is “fast and cheap.”^[6] Mira Murati's Thinking Machines Lab — a twelve-billion-dollar startup led by the former CTO of OpenAI — has released customization tooling (Tinker) built to work across a portfolio of open-weight model families including Qwen and Llama.^[7]

The magnitude of this shift matters for the argument. Entrepreneurs are drifting toward Chinese open-weight models because the economics are decisively better and the capability gap is no longer prohibitive — not because of ideology or geopolitics. The open-weight track has moved from compromise option to default option for a substantial portion of the workloads that generate the displacement pressure.

3. The Meta reversal — and why Google's Gemma 4 matters

The open-weight trajectory is not monotonic, and a clear-eyed account needs to acknowledge the back-sliding. Meta, which built its initial reputation as an open-weight champion through the Llama series, reversed course in late 2025 after its flagship Llama 4 Behemoth model was indefinitely delayed following disappointing internal benchmarks.^[8] Its Superintelligence Lab shifted to developing proprietary successors — models code-named Mango and Avocado targeted for 2026 — and Meta's 2025 capital-expenditure guidance rose to roughly seventy-two billion dollars, with the closed-garden buildout a major component. The Muse Spark launch earlier this month is the first visible artifact of the pivot. The Meta reversal disrupts the simple narrative that frontier labs are converging on open weights. They are not all converging. Some are pulling in the opposite direction, citing the same safety, capability-race, and competitive-economics arguments that Anthropic and OpenAI have long made. The Western open-weight track, once underwritten substantially by Meta's Llama releases, now depends more heavily on Google's willingness to maintain the Gemma line and on the continued vitality of the Chinese open-weight ecosystem.

That said, the one major Western frontier lab that has gone decisively in the open-weight direction is Google. In early April 2026, Google released Gemma 4 under an Apache 2.0 license — multiple model variants ranging from 2B parameters suitable for edge deployment to a 31B dense model that ranks among the top open models on the Arena leaderboard, outcompeting rivals many times its size.Google Open Source^[9] Gemma 4 is built from the same research and technology as Gemini 3; it handles more than 140 languages, supports native function calling, and accepts video and audio inputs. It runs on Hugging Face, Kaggle, and Ollama, and it can operate on hardware from smartphones to consumer GPUs.

Gemma 4 is the counter-example the argument needs. When a Western frontier lab ships Apache 2.0 open weights at near-frontier capability, it does what no regulatory intervention, industry manifesto, or startup coalition could do on its own: it creates the structural latitude entrepreneurs need to avoid platform dependence without giving up capability. Google has effectively voted with its release decisions that a healthy AI ecosystem requires an open-weight track alongside the closed track. This is the correct posture, and it deserves public credit. It also deserves defense, because the Meta reversal shows how quickly open-weight commitments can unravel when internal competitive pressures change.

The frontier labs face their own constraints

Frontier labs are not monolithic winners of the AI race. They face structural constraints of their own. Data-center power demand has outstripped grid capacity in the US electricity markets where frontier training infrastructure concentrates; the GRID Act introduced in early 2026 would require new data centers above twenty megawatts to generate off-grid power.^[10] Capital-expenditure requirements at the frontier are climbing past one hundred billion dollars annually across major labs and hyperscalers combined. Revenue growth has been strong but has not yet demonstrated it can support the infrastructure burn rate sustainably. The alignment-tax structures some labs have built — Anthropic's Project Glasswing, which provides up to one hundred million dollars in usage credits to a consortium of critical-infrastructure firms, is the canonical example — are visible accommodations to capabilities the labs themselves have documented as concerning.^[11] The labs are building extraordinarily expensive infrastructure against a revenue profile still under construction, while simultaneously navigating safety concerns they themselves have disclosed. Their position is strong but not structurally unassailable, and the open-weight alternatives exist in part because the closed alternatives face their own compounding pressures.

Concentration of economic gains

PwC's 2026 finding that roughly three-quarters of AI's economic benefits are captured by the top fifth of companies does specific analytical work.PwC^[2] It establishes that capability access does not distribute gains proportionally to capability use. The gains concentrate above the capability layer, at the platform tier and at the tier of firms that have negotiated favorable platform terms. Small entrepreneurs, displaced workers, and new entrants face the platform architecture on terms that are structurally disadvantageous. Architecture, not capability, explains where the surplus lands.

So what? If architecture is the variable, then the entrepreneur's stack choice is not a procurement decision. It is a distributional decision made at the level of the individual venture.

Engaging the Capability Risks Seriously

I have made the case for open weights as a structural alternative, and I want to be careful not to present that case as if the capability risks that Anthropic, OpenAI, and other frontier labs cite in their defense of closed platforms were not real. They are real. A model with documented autonomous vulnerability-discovery capability — Claude Mythos Preview, to take the most recent example — or with emergent concealment behaviors during evaluation can cause structural harms at scale that are qualitatively different from the harms of previous software categories. Centralized platforms can enforce use policies, apply post-deployment safety fine-tuning, and accept accountability for downstream misuse in ways that distributed deployment cannot match. Open-weight models can be fine-tuned to remove safety guardrails, and the removal of guardrails to enable bioweapon synthesis, autonomous cyberattack, or large-scale disinformation campaigns is not a hypothetical concern. It is a realized operational risk that the open-weight ecosystem has not solved. The serious versions of the safety arguments for closed platforms are genuinely serious, and the response to them cannot be dismissive.

What the argument above insists on is a narrower claim. Even granting the safety concerns in full, the platform architecture they justify is not the only architecture available, and the distributional consequences of choosing a closed architecture for the workloads that do not require the safety protections are substantial. Most of the workloads that generate the labor-displacement pressure we are discussing — content generation, summarization, analysis, coding assistance, domain-specific reasoning — do not sit near the dual-use or catastrophic-misuse frontier. They sit in the middle of the capability-and-risk distribution, where the safety case for closed platforms is weak and the distributional case for open weights is strong. The right intellectual move is to distinguish capability domains where centralized enforcement is warranted from capability domains where it is not, and to resist the extension of safety arguments beyond their actual warrant.

This is the wise-pragmatic posture, not the utopian one. Open weights operate as a structural alternative to platform dependence in the domains where the safety case does not override the distributional case, and they compound alongside credible closed-platform safety architectures in the domains where both need to coexist. The real question is which architecture governs which workload, and by what process that allocation is decided.

The Historical Pattern

The distributional question is not new, but the discourse around AI keeps reinventing it as if it were. Every general-purpose technology wave in the past century and a half has produced both a displacement layer and a platform layer, and the platform layer has consistently been where concentrated value capture occurred. The policy pain attributed to the technology has often been, on closer examination, a consequence of the platform layer's market structure rather than the technology's capability.

Electrification displaced an enormous number of skilled artisans — candle makers, ice harvesters, the cobblers whose crafts lost to machine production. But the darkest chapters of the electrification era happened in company towns and monopsony labor markets where the generation and distribution layer concentrated in a small number of hands. Samuel Insull's Middle West Utilities empire and the holding-company tangle whose collapse in 1932 revealed how much of the electrification surplus had been captured at the infrastructure layer — these were platform-layer pathologies, not technology pathologies.^[12] Where workers owned the end-use tools of the electrical age, where small shops could run their own motors and lighting and compressors on the same grid as General Electric, displacement was absorbed through entrepreneurial entry. Where the platform layer concentrated and extracted, displacement became immiseration.

The internet repeated the pattern. The protocol layer was distributed — TCP/IP, HTTP, SMTP are public goods — and the distribution enabled a decade of broad entrepreneurial entry. The platform layer consolidated in the 2010s into Google, Meta, Amazon, and Apple, and consolidation captured most of the value the protocol layer had enabled. Creators on YouTube, sellers on Amazon, developers shipping through Apple's App Store — they experienced platform-mediated value capture at scale. Algorithm changes demonetized creators. Marketplace policies collapsed seller margins. App store terms extracted take rates as high as thirty percent on transactions. The technology made the activity possible. The architecture determined who kept the money.

Cloud computing repeated it again. The elimination of capital-expenditure barriers enabled millions of startups. The cloud itself concentrated in three hyperscalers who now collect the infrastructure margin on every SaaS business in the Western economy.

The pattern is structural, not accidental. John Meacham, whose formulation Karl Weick later elaborated in his writings on organizational wisdom, defined wisdom as “an attitude taken by persons toward the beliefs, values, knowledge, information, abilities, and skills that are held, a tendency to doubt that these are necessarily true or valid and to doubt that they are an exhaustive set of those things that could be known.”^[13] The displacement narrative we have been telling ourselves about AI is not false — but it is incomplete. The capability layer is doing real work in the economy, and the work is substitutive. Yet the narrative has stopped at the capability layer as if the capability layer were the whole story. The platform layer is where the surplus goes.

The technology made the activity possible. The architecture determined who kept the money.

So what? Three times in a century and a half, we have watched the same movie. The AI wave's architectural choices are live in the current frame.

The Stack Is the Strategy

For entrepreneurs, the practical implication is that stack selection is now a first-order strategic decision with distributional consequences for the venture itself.

Two entrepreneurs, same capability, different architecture

Consider two entrepreneurs building AI-augmented service businesses in 2026. The first builds on a centralized stack: ChatGPT Plus, Claude, Midjourney, Canva Pro, Zapier, selected enterprise APIs. She reaches substantial revenue quickly because the tools are polished, the coordination is handled, and the capabilities are frontier-grade. She is also exposed at every layer of her stack to platform risk: pricing changes that compress her margins overnight, model deprecations that invalidate her workflow, policy changes that reclassify her use case, and a competitive environment in which the platforms themselves may decide to offer her exact service as a first-party feature. She has built a real business — on rented infrastructure whose landlord can raise rent unilaterally.

The second entrepreneur builds on an open-weight stack: a local Qwen or DeepSeek deployment for reasoning, a Gemma 4 edge deployment for latency-sensitive tasks, self-hosted workflow tools, open image and audio models, her own integrations. Her time-to-revenue is slower because she absorbs coordination overhead the platforms would otherwise handle. But her marginal costs are fixed rather than per-query. Her policy exposure is limited to what the Apache 2.0 and MIT-style licenses permit — which for the current open-weight generation covers nearly any commercial use. Her deprecation risk is negligible. The Qwen3 model she deployed in 2026 will run unchanged in 2030 if she chooses not to upgrade. Her competitive exposure to the platforms themselves is substantially lower because the platforms cannot reclassify her customer base as their own.

Operating leverage, bidirectional

The fixed-cost structure of the open-weight stack produces a specific long-term financial benefit that the per-query centralized structure does not produce: operating leverage. When the entrepreneur pays fixed compute costs rather than per-query variable costs, she captures the margin expansion that comes with scale. A thousand additional customers on a centralized stack add a thousand multiples of the per-query cost; a thousand additional customers on an open-weight stack add near-zero marginal cost until compute capacity has to be expanded. Over a three-to-five-year venture trajectory, this operating leverage compounds into a substantial margin advantage — provided the venture reaches the scale that utilizes the fixed-cost base. Operating leverage is bidirectional. For ventures with credible paths to scale, the open-weight stack's fixed-cost structure is a compounding advantage the centralized stack cannot match. For ventures uncertain about their scale trajectory, the variable-cost structure of the centralized stack is the safer short-term bet, because fixed costs without the revenue to amortize them turn into drag rather than leverage. The entrepreneur who understands this tradeoff can choose between them deliberately. The entrepreneur who does not understand it chooses one by default — usually the centralized stack because it is the path of least short-term resistance — and absorbs the long-term distributional consequences without having chosen them.

Platform envelopment

A second risk the platform-dependent entrepreneur carries that the open-weight entrepreneur does not is what Eisenmann, Parker, and Van Alstyne (2011) named platform envelopment.^[14] OpenAI's consumer ChatGPT product has already enveloped much of the value that early API-dependent wrapper startups provided. Perplexity-style search features, document-analysis tools, and coding assistants that entrepreneurs built on the OpenAI API now exist as first-party ChatGPT features, often released without notice to the upstream developers who had built the original products. The entrepreneur on the centralized stack is not merely exposed to pricing and policy risk; she is competing against the platform that supplies her core capability, and the platform has her customer data to guide its entry. Open-weight stacks substantially reduce envelopment risk because the platform does not have direct access to the entrepreneur's customer base or commerce flows.

Digi-serfdom and platform-induced equivocality

The Digital Battlegrounds framework that Richard Hunt, Robert Nugent, Joseph Simpson, Maximilian Stallkamp, Esin Bozdag, and I published in the Academy of Management Annals gives specific language for what is at stake.^[15] Platforms are not neutral transactional infrastructure. They are contested battlegrounds where complex power dynamics determine who captures value and who is captured by the platform. The concept of digi-serfdom that we developed in that paper applies with particular precision to the entrepreneur who builds a business entirely on platforms she does not control. She is structurally bound to the platform's pricing, policy, and strategic decisions in ways that limit her strategic autonomy even as she ostensibly operates her own venture. Rentier capitalism at the platform layer is not a rhetorical provocation; it is the operating condition of a growing share of the AI-augmented venture economy. Open weights are one mechanism for escaping digi-serfdom — not the only mechanism, and not a guarantee of escape, because the compute layer, the regulatory layer, and the distribution layer can each reintroduce platform dependence — but they reduce the platform's capacity to act as feudal lord over the entrepreneur's venture.

The knowledge-problems framework sharpens the point further. The knowledge problems that matter for entrepreneurial action — uncertainty, equivocality, ambiguity, complexity — do not operate independently of the platform architecture. Platform dependence generates a specific species of equivocality: the entrepreneur cannot reliably predict what the platform will do next. Will the pricing change? Will the model deprecate? Will the policy reclassify her use case? Will the platform envelop her product? This is equivocality in the technical sense — multiple plausible futures compatible with the information available — and it is introduced into the venture by the platform layer, not by the capability layer. Open weights do not eliminate knowledge problems. They eliminate the platform-induced equivocality that centralized deployment imports. For the entrepreneur navigating genuine uncertainty about her market, her customers, and her venture's trajectory, removing platform-induced equivocality is one of the few structural parameters she can actually control.

So what? Stack choice is an architectural decision whose effects compound over the lifetime of the venture. Ventures that default into one stack without understanding the tradeoff inherit its distributional consequences whether they wanted them or not.

Open Weights Are Labor Policy

If the analysis above is correct, then the availability of open-weight models at near-frontier capability is not merely a technical question about AI development. It is a structural parameter on the labor-market transition this wave is producing. Open-weight availability determines whether workers displaced from specific task categories have a credible entrepreneurial entry pathway, or whether they face a pure wage loss without a substitution opportunity. It determines whether the savings from capability displacement distribute across the economy or concentrate at the platform layer. It determines the bargaining power of small firms and new entrants relative to the platforms they depend on.

This reframing produces three consequences.

First, regulations that disadvantage open-weight models — compute thresholds that trigger reporting obligations only for open releases, audit requirements that effectively require centralized institutional infrastructure, licensing regimes that gate deployment behind platform-level compliance — are labor-market interventions even when they are framed as safety interventions. Their labor distributional effect belongs on the table alongside the safety considerations. This does not mean the safety considerations should be discounted. It means the distributional consequences should be visible when the safety tradeoff is being made. The EU AI Act's August 2026 high-risk-systems deadline will be the first major test of whether regulatory frameworks can hold both considerations simultaneously.^[16] I am not optimistic that they can. I hope to be wrong.

Second, antitrust analysis of AI platform concentration needs to incorporate the labor pathway. The consumer-welfare standard that has dominated antitrust thinking since the 1980s focuses on end-user pricing and product quality. It is structurally blind to the distributional question of where the surplus from capability-driven productivity gains accrues. The AI wave is generating enormous surplus and concentrating most of it at the platform layer. A framework that cannot see this is not adequate to this moment.

Third, the entrepreneurship-support ecosystem — accelerators, incubators, economic development programs, university entrepreneurship centers — should treat open-weight stack literacy as a first-order capability for the founders they support, alongside genuine frameworks for evaluating the tradeoff. Entrepreneurs need more than access to open-weight models; they need tools and insights for assessing when the capability gap warrants paying the platform-dependence cost and when it does not, how to calibrate operating-leverage tradeoffs against utilization risk, how to recognize envelopment threats before the enveloping firm is already executing the entry, and how to build venture architectures resilient to the specific varieties of equivocality that different platform layers introduce. The short-term acceleration offered by centralized stacks comes bundled with long-term platform dependence that can foreclose strategic options the venture will need as it scales. This is not an argument against centralized stacks in all cases. It is an argument for informed stack choice with the distributional consequences, the operating-leverage tradeoffs, and the envelopment risks made visible to the entrepreneurs whose lives and ventures depend on the choice.

What to Watch

Frank Knight, writing in the years leading to his 1921 monograph, observed that “little or nothing is really fixed but all is a perpetual flux. That which seemed permanent when superficially viewed is seen as the result or product of indefinite transformations.”^[17] The platform architecture of the AI economy looks permanent from inside the current news cycle. It is not. Three developments over the next twelve months will shape whether the argument above proves out empirically.

The first is the trajectory of open-weight capability relative to the centralized frontier. If DeepSeek's next release, Qwen3, the Gemma 4 family, and the next generation of credible open-weight releases continue to close the gap, the strategic case for open stacks strengthens sharply. If Meta's closed pivot signals a broader Western frontier retreat from open weights, or if the Chinese open-weight ecosystem slows, the capability cost of open stacks rises and the tradeoff shifts. The Stanford AI Index plateau and the rapid rise of Chinese open-weight deployment suggest the first trajectory is more likely than the second. But Knight's point holds: the trajectory is not fixed. The Meta reversal is evidence of that. What looked like a consolidating open-weight coalition twelve months ago now depends on a smaller set of committed firms — Google most prominently on the Western side, the Chinese open-weight labs on the other — and the durability of those commitments cannot be taken for granted.

The second is the first major platform-deprecation or envelopment event that visibly destroys downstream businesses. This has happened at small scale already — GPT-3 API deprecations, various Claude model retirements, OpenAI's absorption of features that API-dependent startups had been offering — but the first time a major platform shift collapses a significant portion of the venture ecosystem built on top of it will crystallize the platform-dependence and envelopment risks in ways that data points alone have not. When that event occurs — and I suspect it will occur within the next eighteen months — the strategic calculus for entrepreneurs will update sharply.

The third is the regulatory response. Whether the EU AI Act's enforcement window, US AI governance proposals, and the ongoing debates about open-weight restrictions treat the labor dimension and the distributional consequences as first-order considerations or as background conditions will shape the distributional future of the AI economy for a decade. The precedent that Google's Gemma 4 release establishes — a major Western frontier lab treating open weights as a responsibility rather than a retreat, even as Meta abandons its prior commitments — deserves to be studied, imitated, and defended. The precedent the Chinese open-weight ecosystem establishes — that capability at near-frontier levels can be distributed broadly rather than captured narrowly — deserves the same attention, even as Western regulatory frameworks figure out how to accommodate it.

The displacement narrative that has dominated public discourse since late 2022 has captured something real. AI is substituting for labor at scale, and the scale is accelerating. But the narrative has obscured something structural. The displacement is not the whole story. Where the savings end up, who captures the surplus, whether displaced workers have credible entrepreneurial pathways to value creation using the same capabilities that displaced them — these are distributional questions, not technology questions. They will be answered by the platform architecture underneath the capability, and by the choices — entrepreneurial, regulatory, political — that shape that architecture over the coming years.

The platform is not the capability. The capability can be deployed many ways.

Wise pragmatism, the posture I have been arguing for in almost everything I have written about AI and entrepreneurship, requires holding both the capability story and the platform story at the same time — acting without excessive confidence that the current platform architecture is fixed, and without excessive caution that entrepreneurs cannot change it. The Chinese open-weight ecosystem and Google's Gemma 4 are the early evidence that the architecture is still in motion. The Meta reversal is evidence that the motion can run in both directions. What entrepreneurs need now is tools and insights for navigating the tradeoff honestly — capability gap against platform dependence, short-term velocity against long-term operating leverage, coordination overhead against envelopment risk. Giving them those tools is work the entrepreneurship field can do. Doing that work is what this moment requires.