America’s real AI advantage is in ecosystems
By Michael Frank
Executive summary
American AI policy should:
- Measure adoption, not widgets. Share of workloads, SDK usage, and governed regions are better metrics for American foreign policy goals than chip compute.
- Ban the frontier. Keep truly frontier AI capabilities out of China.
- Win the rest. Maximize lawful China reliance on American ecosystems below the frontier.
American AI policy must end a myopic fixation with chips. Ecosystem dominance is the better goal. Ecosystems are the way to succeed in business, and the advantage that accrues to businesses can also accrue to America’s national interest.
If you care about national power, you should care less about who has access to the fastest chip or best model and more about who sets the defaults—the SDKs, clouds, compatibility regimes, safety gates, and distribution rails that everyone else builds on. Through its companies’ control of ecosystems, America gains leverage that can be useful in a crisis scenario.
“Developers! Developers! Developers! DEVELOPERS! DEVELOPERS!”
At the height of the dot-com boom, Steve Ballmer got on stage at a Microsoft conference with a mission. Ballmer was deeply concerned about the company’s continued prospects. Microsoft was mired in a years-long antitrust lawsuit. So Ballmer—sweaty, exasperated—used that speech to try to communicate what he thought was the only way to make Microsoft’s next thirty years as spectacular as its first thirty. His message was unequivocal: “The key to dot net, the key to industry transformation, the key to success is developers.”[i] He then chanted “Developers!” thirteen more times, clapping aggressively at the audience to get their buy-in.
Ecosystems—not discrete products—decide who sets the rules. It’s why Ballmer was so obsessed with Microsoft becoming an ecosystem company, and it’s why American national security should optimize for ownership of the rails rather than possession of the widget. In almost all of Ballmer’s public speeches in this era, he talks about the importance of the developer ecosystem.[ii] In a recent appearance on the Acquired podcast, Ballmer explained his thinking about why developers were so important—they are the backbone of an ecosystem.[iii] “It’s clear in all these competitions, the thing you need are third-parties that reinforce what you’ve got… I go back to, ‘developers, developers, developers’. I’m trying to tell people at that time that third-parties really mattered.”
Ballmer’s bet on the ecosystem, rather than first-party apps, not only helped Microsoft emerge from the wreckage of the dot-com crash but built the foundation for Microsoft’s incredible bull run since then. The ecosystem enabled Microsoft to capture far more value than they could have in its first-party form through the 1990s. As of today, it is the world’s second most valuable company, with a market capitalization of more than $3.8 trillion.[iv]
The top six most valuable companies in the world—Nvidia, Microsoft, Apple, Alphabet, Amazon and Meta—have all invested mightily in cultivating ecosystems. NVIDIA binds buyers to its hardware through CUDA and its associated libraries—a software stack used by millions of developers that standardizes performance and portability—and, in practice, keeps model training and inference on NVIDIA GPUs.[v] Microsoft combines Azure’s global footprint with a massive organization partner network and over 100 million developers on GitHub, turning distribution and developer gravity into switching costs.[vi],[vii],[viii] Apple monetizes an installed base of over 2 billion active devices through high‑margin services and an App Store that facilitated $1.3 trillion in developer billings and sales in 2024.[ix],[x] Alphabet leverages Android (≈72% of global mobile OS share as of July 2025) and GMS/Play governance to secure default distribution for Search, Maps, and YouTube.[xi],[xii] Amazon runs two classic ecosystems—AWS (140,000 partner organizations worldwide) and the Marketplace, where more than 60% of store sales come from third‑party sellers.[xiii],[xiv] Meta converts 3.5 billion daily active users into advertiser demand while shaping AI tooling via PyTorch under Linux Foundation governance—expanding Meta’s influence over the machine learning supply chain even when it’s not directly monetized.[xv],[xvi]
AI ecosystems have powerful lock-in effects
So, what does it mean to build and own an AI ecosystem? It might be easiest to understand by approaching it from the perspective of companies that sign up for the ecosystem. These decisions are made very early in the lifespan of an AI company or application. The lock-in power is strong. As an applied AI company, the main factors we sought in an ecosystem were cutting-edge models, compute, search, storage, security, and non-AI automation tools (e.g. automated communications). For our purposes, we narrowed the choice to Google and Microsoft, and we chose the latter in large part because of their partnership with OpenAI—which, at the time, seemed to be far ahead of Google with their frontier models. (Today I am not so sure).
Big corporations have existing relationships with ecosystem providers, which can be enormously sticky. Corporations dread technological transformation. There is a heuristic in the B2B SaaS industry that an alternative product needs to be at least ten times better value than an incumbent to induce a company to change. That benchmark is much higher for the entire ecosystem.
Both Microsoft and Google offer hundreds of thousands of dollars in compute credits to start-ups.[xvii][xviii] They are probably the single largest source non-dilutive capital (albeit compute capital) in the world. The credits are not available all at once. Start-ups need to use all the credits in progressively larger tranches before unlocking more. The only way to use the credits is to build within the ecosystem. When a start-up uses all the credits, they will either fold or turn into fast growing customers. At that point, the incentives to stay with the same provider are very strong. The start-up’s entire architecture is dependent on the chosen ecosystem. It would be enormously disruptive to leave.
An ecosystem is a virtuous cycle in a way a product is not. A product, like a chip or a model, can win or lose quickly. An ecosystem is the broader network of complementors, governance, and distribution rails that spans layers. Ecosystems create multi‑node lock‑in with:
- Providing APIs, libraries, tooling
- Distribution defaults (preinstalls, app stores, marketplaces)
- Data/compute billing, storage locality, model tuning)
- Governance (compatibility programs, foundations)
Ecosystems are hard to unwind because customers, developers, hardware vendors, and resellers all become interdependent. Clients commit to the ecosystem. As time goes on, they only become more committed to the ecosystem. The ecosystem owner, meanwhile, is either collecting rents on the ecosystem (like Microsoft Azure and Apple’s App Store); using the ecosystem to lock in other revenue streams (like Nvidia’s CUDA and Google Android); or steering development onto their preferred frameworks that indirectly reinforce revenue (like Meta’s PyTorch).
Ecosystem lock-in can benefit the national interest
When American firms steward the compatibility layer and the distribution rails, APIs, safety and privacy defaults become global norms. This is how standards power operates in practice: through default settings in ecosystems. Android is a good example of that mechanism at play: an open core with a Compatibility Definition Document (CDD) and CTS test suite that preserve uniformity so the services can scale.[xix]
States can exploit chokepoints in global networks to coerce or constrain adversaries—weaponized interdependence.[xx] The problem with the current product-obsessed chokepoint paradigm is that degradation of one product can usually be offset by gains elsewhere in the ecosystem. Product bans accelerate Chinese self‑reliance and cede the pressure points America needs in a crisis scenario. America’s goal should not be autarky. Instead, policy should actively cultivate China’s continued (and deepening) dependency on American tech ecosystems. That dependency can serve as a deterrent to taking actions hostile to American interests, and will actively degrade China’s warfighting capability if deterrence fails.
For AI, that means American ecosystems must directly and aggressively compete with Chinese competitors on all aspects of the AI stack: chips, models, talent, products. It should be the explicit policy of the administration to aggressively target Chinese ecosystem providers’ market share internationally and in China. Nvidia should not only be allowed but encouraged to sell a more reliable and performant chip than Huawei can ship to the domestic Chinese market (while staying below the frontier). OpenAI’s recent release of GPT-OSS is a national imperative to compete against China’s burgeoning leadership in open weight foundation models—models are not a chokepoint since there will always be open source alternatives at or near the frontier. Rather than poaching talent from one another, American companies and universities should aggressively court Chinese AI researchers and bring them to their campuses in the United States.[xxi] Product licenses should include terms that pull China-based users deeper onto American software stacks.
At the same time, it makes sense to continue blocking China’s access to the frontier where China does not have domestic alternatives. That will require companies’ forthright cooperation to establish exactly where a dynamic Chinese AI frontier lies, and make routine adjustments to stay ahead, but only enough to win market share. The Chinese market should not receive priority over domestic American customers. Companies must understand that in the event of a US-China crisis scenario, they must take immediate steps to cut off their customers no matter the cost—similar to what many companies did in Russia after its 2022 invasion of Ukraine. The effect should be ecosystem-wide, not limited to products, since a country—not a company—is the unit of analysis for national security threats. A sledgehammer is better deterrence than a scalpel. Schelling underscores that effective deterrence often rests on blunt, unequivocal threats because their clarity and severity enhance credibility.[xxii]
With respect to ecosystems, policy has better incentive alignment and fewer trade-offs
The legal authorities for these steps are well-established. The legal authorities and triggers in IEEPA are designed specifically for scenarios like a Chinese invasion of Taiwan.[xxiii] There should be clear, auditable triggers. clear notice to allies and partners, and sunset reviews. As part of ongoing trade negotiations, the administration should prioritize American AI ecosystem adoption in competitive global markets like the European Union, India, Singapore, Vietnam, Indonesia, the Philippines, Thailand, Kenya, Nigeria, South Africa, Mexico, and Brazil (similar to the partnership examples of the UAE and Saudi Arabia AI deals).[xxiv] Ecosystem providers need to provision local compute availability, compliance templates, and credits in these markets.
America should explicitly articulate this policy to China’s government and businesses. That communication must take place in private to allow Chinese leadership to save face while avoiding triggers. Recent high-profile communications that America seeks to get China “addicted” to our technology, while correct, undermine the strategy.[xxv]
Of course, China could just implement new policies that block American companies from the market and is moving rapidly in that direction. In many cases, Chinese companies would still find ways to work with American ecosystems. Why? Because they want to use the best possible technology. China’s cutthroat tech ecosystem is filled with companies that don’t care about Communist Party goals; they simply want to win the competition they are engaged in with their Chinese peers. Superior technology is a no-brainer for them. If they are mandated to use Chinese technology, many will still find ways to use American technology. The ingenuity of Chinese technology businesses is limitless. And American firms can shift to prioritize China-adjacent markets and third-country affiliates, preserving leverage in supply chains feeding China-based firms.
Many companies that would be able to take advantage of this policy have repeatedly emphasized the importance of revenue from the Chinese market for R&D and building out domestic manufacturing capability.[xxvi] Let them codify those commitments. The recent announcement of a 15% revenue-share license condition for Nvidia and AMD chip sales to China could be waived and returned to the companies in the future if a greater sum found its way into domestic manufacturing and research.
Unlike the Biden-era export control policy, with its arbitrary thresholds and moving goalposts, this policy should subject itself to verifiable metrics of its efficacy:
- Workload share: percentage of global AI training and inference that takes place on American hyperscalers. (Target: 80% market share, up from ~65% today).[xxvii]
- Framework gravity: share of top‑tier models and toolchains built on American frameworks and compilers (Target: all non-Chinese and more than half of the leading Chinese model developers).
- Governance reach: number of countries operating with American cloud regions, residency, and confidential‑compute enabled (Target: 100% coverage of priority regions by all of the top 5 American cloud providers).
- Dependency in China‑adjacent markets: lawful uptake of American SDKs and clouds below the frontier for non-AI companies. (Target: greater than 90% market share for non-Chinese firms and more than half of Chinese firms).
- Crisis readiness: rehearsed playbooks for ecosystem‑level cutoff including access revocation, policy flips, and coordinated export holds. (Target: complete cutoff within 24 hours of an IEEPA determination across all affected companies).
[i] “Steve Ballmer at NET Conference Going Crazy about Developers! | 1999,” YouTube, November 26, 2023, https://www.youtube.com/watch?v=8fcSviC7cRM.
[ii] Steve Ballmer, “Archives: Speeches,” Microsoft, accessed August 20, 2025, https://news.microsoft.com/speeches/page/94/.
[iii] Steve Ballmer, Ben Gilbert, and David Rosenthal, “The Steve Ballmer Interview,” Acquired podcast, June 1, 2025, https://www.acquired.fm/episodes/the-steve-ballmer-interview.
[iv] Companies Ranked by Market Cap, accessed August 20, 2025, https://companiesmarketcap.com/.
[v] Stephen Nellis, Reuters, March 25, 2024, https://www.reuters.com/technology/behind-plot-break-nvidias-grip-ai-by-targeting-software-2024-03-25/; “NVIDIA CUDA‑X Libraries,” NVIDIA, https://developer.nvidia.com/gpu-accelerated-libraries; “CUDA‑X,” NVIDIA, https://www.nvidia.com/en-us/technologies/cuda-x/.
[vi] Rodney Clark, “Evolving Microsoft Partner Network Programs for Partner Growth and Customer Success,” Official Microsoft Blog, March 17, 2022, https://blogs.microsoft.com/blog/2022/03/16/evolving-microsoft-partner-network-programs-for-partner-growth-and-customer-success/.
[vii] “Azure Regions List,” Microsoft Learn, July 31, 2025, https://learn.microsoft.com/en-us/azure/reliability/regions-list.
[viii] Thomas Dohmke, “100 Million Developers and Counting,” GitHub Blog, January 25, 2023, https://github.blog/news-insights/company-news/100-million-developers-and-counting/.
[ix] “Apple Reports First Quarter Results,” Apple Newsroom, May 8, 2025, https://www.apple.com/newsroom/2024/02/apple-reports-first-quarter-results/.
[x] “Global App Store Helps Developers Reach New Heights, Supporting $1.3 Trillion in Billings and Sales in 2024,” Apple Newsroom, June 5, 2025, https://www.apple.com/in/newsroom/2025/06/global-app-store-helps-developers-reach-new-heights/.
[xi] “Mobile Operating System Market Share Worldwide,” StatCounter Global Stats, accessed August 20, 2025, https://gs.statcounter.com/os-market-share/mobile/worldwide.
[xii] “Android Compatibility Definition Document,” Android Open Source Project, May 29, 2025, https://source.android.com/docs/compatibility/cdd.
[xiii] “Say Hello to 268 New AWS Competency, Service Delivery, Service Ready, and MSP Partners Added in June,” July 3, 2025, https://aws.amazon.com/blogs/apn/say-hello-to-165-new-aws-competency-service-delivery-service-ready-and-msp-partners-added-in-august/.
[xiv] “Amazon Selling Stats,” Amazon, July 14, 2025, https://sell.amazon.com/blog/amazon-stats.
[xv] “Meta Reports Second Quarter 2025 Results,” Meta, July 30, 2025, https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Reports-Second-Quarter-2025-Results/default.aspx.
[xvi] “PyTorch Strengthens Its Governance by Joining the Linux Foundation,” PyTorch, November 15, 2024, https://pytorch.org/blog/pytorchfoundation/.
[xvii] “Start Building with Microsoft,” Microsoft for Startups, accessed August 20, 2025, https://portal.startups.microsoft.com/signup.
[xviii] “Start and Scale with the Google for Startups Cloud Program,” Google Cloud, accessed August 20, 2025, https://cloud.google.com/startup/apply.
[xix] “Android Compatibility Definition Document,” Android Open Source Project, May 29, 2025, https://source.android.com/docs/compatibility/cdd.
[xx] Henry Farrell and Abraham L. Newman, “Weaponized Interdependence: How Global Economic Networks Shape State Coercion,” International Security 44, no. 1 (July 1, 2019): 42–79, https://doi.org/10.1162/isec_a_00351.
[xxi] Mike Isaac, Eli Tan, and Cade Metz, “A.I. Researchers Are Negotiating $250 Million Pay Packages. Just Like NBA Stars.,” The New York Times, July 31, 2025, https://www.nytimes.com/2025/07/31/technology/ai-researchers-nba-stars.html.
[xxii] Thomas Schelling, “An Essay on Bargaining,” essay, in The Strategy of Conflict (Harvard University Press, 1960).
[xxiii] “50 U.S. Code § 1701 – Unusual and Extraordinary Threat; Declaration of National Emergency; Exercise of Presidential Authorities,” Legal Information Institute, accessed August 21, 2025, https://www.law.cornell.edu/uscode/text/50/1701.
[xxiv] Dylan Patel et al., “AI Arrives In The Middle East: US Strikes A Deal with UAE and KSA,” SemiAnalysis, May 16, 2025, https://semianalysis.com/2025/05/16/ai-arrives-in-the-middle-east-us-strikes-a-deal-with-uae-and-ksa/.
[xxv] Zijing Wu and Cheng Leng, “Beijing Turns against Nvidia’s AI Chip after ‘insulting’ Lutnick Remarks,” Financial Times, August 21, 2025, https://www.ft.com/content/b8e30c54-b71c-4113-8b3e-8f54bc36587d.
[xxvi] Alex Irwin-Hunt, “Nvidia and AMD Deals Highlight the US’s Chip Conundrum,” FDI Intelligence, August 13, 2025, https://www.fdiintelligence.com/content/2c00400a-1af7-4b07-aa3f-efb278ee6c2d.
[xxvii] “Worldwide Cloud Service Spending to Grow by 19% in 2025,” Canalys Newsroom, February 20, 2025, https://www.canalys.com/newsroom/worldwide-cloud-service-q4-2024.