OpenAI solves 80-year math problem | Intel undercuts Nvidia

In today's issue

01 · Models

OpenAI model disproves Erdős unit distance conjecture after 80 years

OpenAI's internal AI model has disproven the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry that asked how many pairs of points can sit exactly one unit apart as n grows large. Fields Medalist Tim Gowers called it a milestone in AI mathematics—the first time an AI system has produced a full proof resolving a major open conjecture.

The details:

The proof uses a grid with spacing 1/√65, where each interior point has 16 unit-distance neighbors, leveraging the fact that both 1² + 8² and 4² + 7² equal 65.
Erdős predicted the maximum number of unit distances grows as n^(1+o(1)), but the AI showed the true bound sits higher, disproving his 80-year-old conjecture.
The AI combined existing techniques from graph theory, number theory, and discrete geometry rather than inventing new mathematical methods; human mathematicians have since cleaned up and extended the result.

Why it matters: This is the clearest signal yet that AI has moved beyond pattern matching into genuine mathematical discovery—not by inventing new theory, but by assembling known tools in ways humans hadn't. The real win isn't the proof itself; it's that a major open conjecture fell to a system that combines subfields intelligently. Watch for similar breakthroughs in constraint-heavy domains where brute-force exploration of existing frameworks pays off.

Read the full story →

02 · Analysis

US AI economy is growing 2,000% a year, hidden inside GDP statistics

A new paper from economists at UVA, Anthropic, and the Bank of Canada found the US AI economy hit $250 billion in 2025 and is growing at 2,600% annually in quality-adjusted terms — yet barely shows up in official GDP figures. The gap matters because finance ministries using conventional data will dangerously underestimate the risk of AI-driven labor displacement and tax-base collapse.

The details:

US AI compute spending exploded from $37 billion in 2023 to $219 billion in 2025, while computing capacity more than doubled each year.
Quality-adjusted AI output grew 2,290% in 2024 and 2,271% in 2025, but nominal revenue stayed flat because per-unit prices fell almost as fast as capability rose.
Authors recommend statistical agencies build dedicated AI satellite accounts to track compute spending, training-versus-inference splits, and quality-adjusted output as permanent data series.

Why it matters: This paper is the clearest signal yet that AI is economically massive and statistically invisible — a dangerous combination for policymakers. Unlike semiconductors and the internet, AI has a credible substitution path for human labor, which means governments flying blind on real AI output face a fiscal cliff they won't see coming. The measurement problem is now a policy emergency.

Read the full story →

03 · News

Nvidia unveils RTX Spark, a Windows PC class built for local AI agents

Nvidia launched RTX Spark, a Windows PC built to run AI agents locally without sending data to the cloud. The machine packs 1 petaflop of compute and 128GB memory, ships this fall, and includes OpenShell—a runtime that lets users control what agents can access and which tasks stay private versus go to cloud models.

The details:

Llama.cpp multi-token prediction hits 2x throughput on Qwen3.6-27B and 1.6x on Qwen3.6-35B models running on GeForce RTX 5090.
Nvidia and Microsoft co-developed Windows security primitives for agents, handling identity, containment, and policy enforcement across local and cloud workloads.
Adobe is rebuilding Photoshop and Premiere for RTX Spark, promising 2x faster AI, editing, coloring and effects operations.

Why it matters: Nvidia is betting that privacy-conscious users and creative professionals will pay for local AI inference instead of cloud APIs. The RTX Spark ecosystem—bundled hardware, optimized models, and strict containment rules—turns the PC back into a trusted compute boundary. If Adobe's rebuilds deliver on speed, this could be the first real competitive threat to cloud-only AI workflows.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Models

Intel's Crescent Island AI chip ships this year, undercutting Nvidia on cost and cooling

Intel is shipping its Crescent Island AI chip in limited quantities by year-end, undercutting Nvidia and AMD by using cheaper LPDDR5 memory and air cooling instead of expensive HBM and liquid-cooling systems. This is Intel's first major AI product under new CEO Lip-Bu Tan and targets the inference market—where memory bandwidth is looser and price sensitivity higher than in training, where Nvidia dominates.

The details:

Crescent Island took 18 months to develop and uses air cooling plus LPDDR5 memory, cutting both bill-of-materials and data center infrastructure retrofit costs that Nvidia's Blackwell demands.
Intel's data center chief Kevork Kechichian, who joined from Arm last year, said the company could sell certain chip tiers in China under US export controls—a revenue door largely closed to Nvidia and AMD.
Intel shares have surged over 200 percent since the start of this year, and the US government took a 10 percent stake in August to keep Intel's foundry operations domestic.

Why it matters: Intel is making a smart tactical bet: cede training dominance to Nvidia and fish in the inference pond where margin-conscious customers hate liquid cooling and HBM pricing. If yields hold and Crescent Island ships on time, it could grab real traction in edge inference and cost-sensitive cloud deployments—segments Nvidia doesn't obsess over. The China angle is gravy. Watch whether larger reasoning models make LPDDR5 obsolete faster than Intel expects.

Read the full story →

05 · Business

France lands $108B in foreign investment, half tied to SoftBank AI data centers

France announced $108 billion in foreign investment commitments, with $54 billion—half the total—earmarked for SoftBank's AI data center buildout. This transforms the headline from diversified capital attraction into a strategic bet: France is positioning itself as Europe's primary hub for AI compute infrastructure.

The details:

SoftBank's $54 billion commitment aligns with its earlier €75 billion pledge to build 5GW of AI data center capacity in France, the electricity equivalent of a mid-sized European country's industrial sector.
France's 70% nuclear-powered electricity grid and historically lower industrial power prices than Germany and the UK give it a structural cost advantage competitors cannot easily replicate.
The remaining $54 billion covers manufacturing, pharma, and tech deals from other foreign companies, but SoftBank's single infrastructure bet dwarfs every other line item in the package.

Why it matters: SoftBank is positioning itself as the landlord for the AI economy—owning compute capacity that frontier labs lease regardless of who wins the model race. France just won the location lottery, but the real winner is SoftBank, which now anchors European infrastructure while backing OpenAI's Stargate in the U.S. and building across Asia. Watch for German and Nordic governments to scramble with competing incentives.

Read the full story →