Two Scenarios for 2028: Anthropic Sketches a 12 to 24 Month AI Lead, or a Neck-and-Neck Race

Hello, it's Shii-chan!

Anthropic News anthropic.com

Today I'm covering a policy paper on AI leadership in 2028 from Anthropic's News page. It's written from Anthropic's point of view, so keep that in mind as you read.

What was announced?

Anthropic published "2028: Two scenarios for global AI leadership," a policy paper on US-China AI competition describing two possible 2028 outcomes.

Its core argument is that democracies, not authoritarian regimes, should lead AI development, because the leading developers will shape the rules and norms for how AI is built and used.

Why it matters

Anthropic sees compute (advanced chips) as the most important ingredient for frontier AI.

• Democracies currently hold a substantial compute lead
• Chinese labs have strong talent but are limited by compute
• Two workarounds help them stay close:
  • Illicit or evasive compute access, such as chip smuggling and offshore data centers
  • "Distillation attacks," where fraudulent accounts systematically extract outputs from US models

The paper estimates Huawei will produce about 4% of NVIDIA's aggregate compute in 2026 and 2% in 2027, while acknowledging that such estimates are uncertain.

What changes: the two scenarios

Scenario 1: the US and allies hold a commanding lead

• Export controls tighten, loopholes close, and distillation is curbed
• US models are 12 to 24 months ahead and the lead grows; China wouldn't reach comparable capabilities until 2029 or 2030
• American AI becomes the backbone of the global economy, and democracies set the norms

Scenario 2: the CCP's ecosystem is neck-and-neck

• Chinese models are only a few months behind, helped by distillation, overseas compute access, weak enforcement, and loosened export controls
• Huawei and Alibaba data centers spread globally, especially in lower-cost markets
• The PLA's AI-enabled cyber capabilities become a serious threat

The paper calls the "race" framing misleading and describes a contest across four fronts: intelligence, domestic adoption, global distribution, and resilience.

Dive Deep

Policy recommendations

• Close compute loopholes: tighten controls on smuggling, foreign data center access, and chipmaking equipment (including DUV servicing), and raise enforcement budgets
• Defend US innovations: clarify in legislation that distillation attacks are illegal, and improve threat intelligence sharing
• Export American AI: promote adoption of trusted US hardware and models abroad

Safety and security concerns

• Mythos Preview, released to select partners through Project Glasswing in April 2026, can autonomously find and chain software vulnerabilities. Firefox fixed more security bugs in one month than in all of 2025
• As of the prior year, only 3 of 13 leading Chinese labs had published any safety evaluation results, and none had published CBRN evaluations
• CAISI found DeepSeek's R1-0528 complied with 94% of overtly malicious requests under a common jailbreak technique, versus 8% for US reference models
• Close competition could pressure labs to release models faster with less safety testing

Caveats

• The scenarios are hypothetical, and the 12 to 24 month lead is a target, not a measured fact
• Capability comparisons rely on estimates
• Many cited claims come from third-party reports that the paper doesn't independently verify

Wrap-up

• Anthropic argues compute controls and anti-distillation measures will shape the 2028 AI landscape
• Scenario 1 is a 12 to 24 month lead; Scenario 2 is a race a few months apart
• Recommendations: close loopholes, make distillation clearly illegal, and export US AI
• It's a policy paper from Anthropic's perspective, so weigh the uncertainty in its estimates

This one is for anyone following AI, security, and policy!