Anthropic's 2028 AI Leadership Essay: Two Scenarios for the US-China AI Race
Anthropic's essay outlines two futures: US maintains compute dominance or China catches up. Here's what the argument means for AI builders and enterprise teams.
What Anthropic Is Actually Arguing About AI’s Next Four Years
Anthropic isn’t known for publishing geopolitical strategy papers. The company’s public output tends to run toward model cards, safety research, and constitutional AI principles. That’s what makes its essay on US-China AI leadership notable — it’s a direct, blunt argument about who controls the future of AI and what happens if that changes by 2028.
The core claim: the decisions governments and companies make in the next few years will determine whether the United States — and by extension, safety-focused AI development — leads the most consequential technology in human history. Claude, Anthropic’s AI model, is central to that argument. But the stakes extend far beyond any single model.
Here’s a breakdown of what Anthropic argues, why 2028 is the inflection point, and what it means for enterprise teams and AI builders who are thinking seriously about where this technology is heading.
The Two Scenarios Anthropic Lays Out
The essay’s structure is built around a binary: either the US maintains a meaningful lead in AI compute by 2028, or it doesn’t. Everything else — safety norms, economic outcomes, global AI governance — flows from which of those two worlds you’re living in.
Scenario One: US Maintains Compute Dominance
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
In this scenario, US export controls on advanced semiconductors hold. China faces continued restrictions on accessing the most capable AI chips — primarily Nvidia’s H100 and H200 series and their successors. American AI labs, operating with frontier hardware, continue to push the capability frontier faster than their Chinese counterparts.
The implications Anthropic draws are significant:
- Safety-focused labs stay at the frontier. If the leading AI models are built by organizations that treat alignment and interpretability as core research priorities, the norms around AI development get set by those organizations.
- Democratic governments have more leverage. US-aligned companies can be subject to oversight, policy frameworks, and international agreements in ways that Chinese state-backed AI programs cannot.
- The economic benefits accrue to open societies. The productivity gains from frontier AI compound for countries and companies with access.
This scenario isn’t presented as guaranteed or easy. It requires active policy choices — sustained export controls, continued CHIPS Act investment, and diplomatic coordination with allies.
Scenario Two: China Catches Up in Compute Access
The second scenario is what Anthropic wants policymakers to take seriously as a risk. If export controls erode — through smuggling, third-party workarounds, or policy rollback — or if China develops competitive domestic chip manufacturing, the compute gap narrows.
The consequences Anthropic outlines:
- Frontier AI is no longer dominated by safety-focused actors. A compute-competitive China means the labs pushing the capability frontier include organizations with different priorities around transparency, alignment research, and external oversight.
- Global AI norms shift. Technical standards, safety practices, and deployment norms get set by whoever is building the most capable systems. If that’s no longer US labs, those norms look different.
- The geopolitical leverage that comes with AI leadership weakens. AI isn’t just a product — it’s infrastructure for military systems, economic planning, scientific research, and surveillance. Who controls the frontier matters for everything downstream.
The essay isn’t predicting this outcome. It’s arguing that treating it as a real possibility should change how urgently the US acts.
Why Compute Is the Lever That Matters
To understand the essay’s argument, you need to understand why Anthropic focuses so heavily on compute rather than talent, data, or software.
AI capability — at least at current scales — is largely a function of how much compute you can throw at training. The largest language models require thousands of the most advanced GPUs running for months. There’s a direct relationship between chip access and what models you can build.
Talent and data are more mobile. Engineers move. Data can be scraped or licensed. But advanced semiconductor manufacturing is geographically concentrated and capital-intensive in ways that make it a real chokepoint.
TSMC in Taiwan manufactures the most advanced chips in the world. Nvidia designs them. Neither is Chinese-owned, and both are subject to US export controls. That’s the lever Anthropic is pointing at.
The argument isn’t that software doesn’t matter — it does. Chinese labs like DeepSeek have shown that algorithmic efficiency can partially compensate for compute constraints. But “partially” is doing a lot of work in that sentence. At the frontier, compute still determines the ceiling.
The 2028 Timeframe: Why It’s Not Arbitrary
The essay’s focus on 2028 is specific for a reason. That’s approximately the window within which several converging factors either lock in or shift:
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Domestic chip manufacturing. CHIPS Act investments in US semiconductor fabrication are expected to come online in the mid-to-late 2020s. TSMC’s Arizona fabs, Intel’s expansion, and Samsung’s US facilities all have timelines in this range. If US-aligned production capacity increases significantly, compute advantage becomes less dependent on export controls alone.
China’s domestic semiconductor progress. Huawei and SMIC have made slower-than-expected progress on advanced node manufacturing, but progress is ongoing. By 2028, the gap could narrow. It could also remain large. The essay is arguing that assuming it remains large without active policy effort is a mistake.
AI capability jumps. If current scaling trends continue — or if architectural breakthroughs change the compute-to-capability ratio — the models of 2028 will be substantially more capable than today’s. Whoever leads that development will have capabilities that matter at a national-security level, not just a commercial one.
The 2028 framing is Anthropic’s way of saying: the window for policy action isn’t indefinite.
What This Means for AI Safety Arguments
Anthropic’s safety positioning is embedded throughout the essay, and it’s worth being direct about what they’re claiming.
The argument isn’t purely nationalistic. Anthropic isn’t saying US AI is good because it’s American. They’re saying US AI development — specifically, development led by labs that take alignment research seriously — is more likely to produce AI systems that are interpretable, controllable, and aligned with human values.
That’s a substantive claim, not a trivial one. It means:
- Safety investment follows incentive structures. Labs competing for top AI talent in a culture that values safety research are more likely to build safety into their development processes.
- Oversight is possible. US-based companies can be regulated, audited, and held accountable in ways that labs operating under a different political system cannot.
- Anthropic has skin in this game. Claude is one of the leading frontier models. Anthropic being at the frontier matters to their safety argument — if safety-focused labs fall behind, the capability frontier gets set by actors with different priorities.
This makes the essay both a policy argument and a competitive one. Anthropic benefits if the argument succeeds. That doesn’t make the argument wrong, but it’s worth keeping in mind when evaluating the framing.
Implications for Enterprise Teams
If you’re running an enterprise AI strategy, the US-China compute race probably feels distant from decisions about which models to deploy for document processing or customer support. But the essay’s arguments have practical implications that matter.
Model availability isn’t guaranteed
The AI model landscape that exists today — Claude, GPT-4o, Gemini, Llama — reflects a specific compute and regulatory environment. If that environment shifts significantly, model availability, pricing, and capability could all change in ways that affect enterprise AI deployments.
Building workflows that depend heavily on a single provider or model family carries more risk than building on infrastructure that can swap models. That’s true regardless of geopolitics.
Regulatory changes will follow
The policy debate Anthropic is engaging isn’t abstract. Export controls on chips affect the cost and availability of cloud compute. AI regulation is coming regardless of who leads it. Enterprise teams that understand what’s driving the policy environment are better positioned to anticipate compliance requirements and infrastructure decisions.
The safety argument has enterprise relevance
Day one: idea. Day one: app.
Not a sprint plan. Not a quarterly OKR. A finished product by end of day.
Anthropic’s push to keep safety-focused labs at the frontier isn’t just a geopolitical argument — it translates into model behavior that matters for enterprise use. Models built with stronger alignment research, interpretability tools, and safety testing behave more predictably in production. For regulated industries, that’s not a nice-to-have.
How AI Builders Can Think About Model Diversity
One practical takeaway from the compute dominance argument is that the AI model landscape will likely remain heterogeneous — multiple frontier labs, multiple architectures, multiple providers. For teams building AI applications, that’s actually good news if you’re set up to take advantage of it.
Building on a platform that isn’t locked to a single model means you can swap in whichever model performs best for a given task, or hedge against changes in availability and pricing. Multi-model flexibility isn’t just a technical convenience — it’s a resilience strategy.
This is exactly where MindStudio is useful for enterprise AI teams. The platform gives you access to 200+ models — including Claude, GPT-4o, Gemini, and others — from a single interface, without needing separate API accounts or managing credential overhead. When Anthropic ships a new Claude version, or when a different model performs better on your specific use case, you can switch without rebuilding your workflows.
For teams that want to build AI agents and automated workflows — whether that’s processing documents, handling customer requests, or running internal automation — MindStudio’s no-code builder lets you prototype in under an hour and deploy across multiple models. You can try it free at mindstudio.ai.
The compute race Anthropic is describing will produce a landscape where the best model for a given task keeps changing. Building on flexible infrastructure is the practical response to that uncertainty.
Criticism and Counterarguments
Anthropic’s essay has drawn some pushback worth engaging with honestly.
The efficiency counterargument. DeepSeek’s R1 and other efficient models have shown that the gap between compute-constrained and compute-abundant labs isn’t as fixed as Anthropic implies. If algorithmic improvements continue to narrow the compute advantage, export controls matter less.
Anthropic’s response to this is essentially: efficiency matters at the margin, but at the frontier, compute still determines the ceiling. A lab with 10x the compute can run 10x more experiments, find better architectures, and train larger models. The efficiency argument doesn’t eliminate the advantage — it just compresses it.
The self-interest problem. Anthropic is a US-based lab arguing that US export controls should remain in place and that safety-focused labs (like Anthropic) should stay at the frontier. The alignment of their policy argument with their commercial interests is obvious. That doesn’t make the argument wrong, but it’s a reason to read it critically.
The international coordination problem. Even if US export controls hold, allies — including South Korea, the Netherlands, Japan, and Taiwan — have their own commercial interests in selling chips. ASML’s EUV machines are Dutch. Coordination isn’t automatic.
The stability argument cuts both ways. Some analysts argue that aggressive US compute restrictions accelerate China’s domestic semiconductor development by removing the incentive to rely on US supply chains. If that’s true, export controls might be self-defeating in the long run.
These aren’t trivial objections. Anthropic’s essay doesn’t engage them in great depth, which is a fair criticism of the document.
What Anthropic Is Asking For
The essay isn’t abstract advocacy. Anthropic is making specific asks of US policymakers:
- Maintain and enforce chip export controls. Close loopholes that allow advanced chips to reach China through third countries.
- Invest in domestic semiconductor manufacturing. The CHIPS Act is a down payment; sustained investment is needed.
- Develop AI safety standards that become international norms. If US labs lead, their practices become the default. That’s an opportunity to embed safety standards globally.
- Engage allies. Coordinating with the EU, Japan, South Korea, and Taiwan on export control enforcement and joint AI safety frameworks.
These are policy-level requests, but enterprise AI teams should track them. Chip supply constraints affect cloud compute prices. International AI standards affect what enterprises can deploy in regulated markets. None of this is background noise.
Frequently Asked Questions
What is Anthropic’s essay about the US-China AI race?
Anthropic published an essay arguing that the United States must maintain its lead in AI compute — primarily through semiconductor export controls and domestic chip investment — to ensure that safety-focused AI labs remain at the frontier. The essay frames 2028 as a critical window, after which catching up becomes significantly harder. It outlines two scenarios: one where the US holds its compute advantage, and one where China closes the gap.
Why does Anthropic focus on 2028 specifically?
2028 is the approximate window when several key factors converge: CHIPS Act semiconductor investments come online, China’s domestic chip manufacturing efforts mature (or don’t), and AI capability jumps significantly from current baselines. Anthropic argues that policy decisions made before 2028 largely determine which competitive scenario the world enters.
What does compute dominance mean in the context of AI?
Compute dominance means having preferential access to the most advanced AI chips — currently Nvidia’s H100/H200 GPUs and their successors — at scale. Training frontier AI models requires thousands of these chips running for weeks or months. Labs with more compute can train larger models, run more experiments, and push capabilities faster than labs with restricted access.
Does Anthropic’s argument address the DeepSeek efficiency concern?
Partially. Anthropic acknowledges that algorithmic efficiency can partially offset compute constraints — DeepSeek demonstrated this by building competitive models with fewer resources. But Anthropic’s position is that efficiency advantages compress but don’t eliminate the compute gap. At the frontier, raw compute still determines the ceiling of what’s achievable.
Is Anthropic’s argument self-serving?
Openly, yes. Anthropic is a US-based lab arguing that US export controls should stay in place and that safety-focused labs should maintain frontier leadership. Their commercial interests align with their policy argument. That’s a legitimate reason to apply scrutiny to the framing, but it doesn’t automatically invalidate the underlying claims about compute, safety, and geopolitical stakes.
How does the US-China AI race affect enterprise AI strategy?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
For enterprise teams, the race matters in a few concrete ways: model availability and pricing are affected by the competitive landscape, regulatory requirements will evolve based on which actors are leading AI development, and the safety practices of frontier labs influence how predictably and safely deployed AI behaves in production. Building on infrastructure that supports multiple models hedges against disruptions in any single part of the supply chain.
Key Takeaways
- Anthropic’s essay frames 2028 as the critical window for US-China AI leadership — after which the trajectory becomes much harder to alter.
- The two scenarios are: the US maintains compute dominance through export controls and domestic chip investment, or China closes the gap and changes who sets the global AI frontier.
- Compute matters because training frontier models requires advanced chips at scale, and that hardware supply chain is currently US-controlled.
- The safety argument is substantive: labs that take alignment research seriously set different norms than those that don’t — and those norms propagate through deployment at scale.
- For enterprise teams, the practical response is building on infrastructure that supports model flexibility, understanding the regulatory trajectory, and preferring AI providers with transparent safety practices.
- MindStudio’s multi-model platform lets teams build workflows that aren’t locked to any single provider — useful in a landscape where the best model will keep changing. Start free at mindstudio.ai.
