The latest Epoch Brief presents several connected findings: a detailed review of Huawei’s AI chip roadmap and production constraints, evidence that US GDP statistics undercount Nvidia’s contribution by roughly 0.3 percentage points, measurements showing architectural differences between OpenAI’s GPT and Anthropic’s Claude on long contexts, and a series of data insights and benchmarks — including a pre-release evaluation of GPT-6 Astra and initial progress on the FrontierMath Erdős benchmark.
Research: Can Huawei catch up to Nvidia by 2030?
Researcher Venkat Somala examined Huawei’s chip roadmap and supply chain — from the Ascend 950 through its 3D-stacking strategy and domestic HBM supply — and finds a steep climb is required for Huawei to compete with industry leader Nvidia. Key points:
- Huawei’s most powerful chip delivers roughly half the arithmetic performance of the Nvidia H100.
- Somala’s estimates indicate Huawei will likely produce less than 4% of Nvidia’s AI compute in 2026.
- Huawei has announced an ambitious roadmap, but the two primary levers to close the gap — improving per-chip performance and increasing shipped volume — face constraints from export controls and supply limitations.
The report concludes that even if Huawei fully executes its roadmap, it is very unlikely to catch Nvidia within this decade.
The Nvidia-sized hole in US GDP statistics
A team led by senior researcher Isabel Juniewicz found that US GDP statistics miss much of the value that American firms — notably Nvidia — create by designing AI chips that are manufactured and sold abroad. Consequences include:
- Official GDP figures understate real GDP growth by about 0.3 percentage points over the past year.
- This helps explain why measured GDP effects appear modest despite US investment in computing equipment rising to roughly $400 billion per year — nearly three times the 2023 level.
- The underlying issue is that domestic design and intellectual value captured by US firms does not fully register in GDP when production and sales occur overseas.
Long-context latency: GPT vs Claude
Jason Li’s analysis of time to first token (TTFT) with context lengths up to one million tokens suggests architectural differences between OpenAI’s GPT-5.6 family and Anthropic’s Claude 5 family:
- GPT models show a substantial quadratic component in latency scaling as context length increases, and costs rise notably past about 272k input tokens.
- Claude models scale closer to linearly, with cost remaining more stable at long context lengths.
These measurements are consistent with differing architectural choices across the families, though the report remains cautious about asserting direct causality.
Data insights
- The record for the largest AI data center by IT power capacity has doubled every ten months since mid-2024. The current record-holder is SpaceXAI’s Colossus 2, at an estimated 950 MW.
- Since the introduction of the first reasoning models (OpenAI’s o1-mini and o1-preview) in September 2024, the Epoch Capabilities Index (ECI) frontier has advanced linearly by 14 points per year; non-reasoning models have progressed at about 6 points per year.
Commentary: revenues and sustainability
Anthropic and OpenAI were already among the fastest-growing companies of their scale in 2025, and their revenue growth accelerated further in 2026. Researchers Josh You and Lynette Bye discuss the implications and sustainability of that acceleration in the latest Gradient Updates newsletter.
Benchmarking: FrontierMath Erdős and GPT-6 Astra
- OpenAI’s GPT-6 Astra (pre-release access used for these evaluations, tied to OpenAI’s September 3 release) set multiple new records across the Epoch benchmarks and took the top spot on the ECI among 247 tracked models.
- On FrontierMath Erdős — a new benchmark of 68 formalized open Erdős problems encoded in Lean — GPT-6 Astra is the only model to make measurable progress so far, solving 2 out of 68 problems.
- Astra scored 98% on FrontierMath: Tier 4, solving the final outstanding problem at that level; the benchmark is now considered saturated at that tier.
Other updates
- The new AI Chip Users explorer compiles estimates of the compute frontier used by leading AI developers. It estimates OpenAI increased its compute 17-fold in two years, a striking example of industry-wide compute growth.
Careers
Epoch is hiring across multiple roles: Researcher/Senior Researcher, Data Scientist, Software Engineer (Benchmarking), Senior Product Designer, Writer & Editor, Social Video Producer, and IT and Security Specialist. Applications are rolling, and there are also options to submit a general expression of interest or apply for Special Projects Associate roles.
ICYMI (selected recent pieces)
- Most AI use at work happens on free plans, except in science and tech (August 14, 2026)
- Performance per dollar of purchased AI chips has grown about 49% per year (August 13, 2026)
- Commentary: nine big questions benchmarks can help answer (August 14, 2026)
- Will financing bottleneck AI compute? An Anthropic case study (August 12, 2026)
Closing
The Brief’s findings highlight three themes: hardware and supply-chain constraints materially shape who leads the AI hardware market; current economic statistics can undercount the macroeconomic value of digital design and intellectual property; and architecture-level differences across model families have practical cost and latency consequences at large context scales. Meanwhile, capability benchmarks and infrastructure metrics continue to advance rapidly, as illustrated by GPT-6 Astra and the FrontierMath Erdős results.
