Claude advances Riemann zeta function lower bound to 67.2% via agentic reasoning

Anthropic locks in $19.1B in structured compute deals with alternative energy providers, signaling shift in AI infrastructure financing

Anthropic has secured two major long-term compute agreements totaling $19.1B: a $9.1B, 20-year deal with bitcoin miner Riot Platforms for 191 MW of capacity at its Rockdale, Texas facility, and a separate $10B six-year infrastructure commitment with a Norwegian bitcoin miner. These deals represent a fundamental shift in how frontier labs finance AI scaling—moving away from traditional capex on their own balance sheets toward structured, long-term partnerships with alternative energy providers who can monetize excess or repurposed mining infrastructure. The pattern mirrors Amazon's 7.65 GW natural gas plant in Texas and Google's $200B TPU financing structures, indicating that AI infrastructure economics are increasingly decoupling from model labs' direct ownership and instead flowing through energy-backed financing vehicles.

Deepseek and Alibaba reach cost-performance parity with US frontier labs, then raise prices—undermining cheap-alternative narrative

Chinese AI labs Deepseek and Alibaba have achieved Pareto frontier performance alongside Meta and Grok, but both have responded by raising prices rather than competing on cost. Deepseek announced significant price increases while Alibaba is exploring revenue-sharing models, contradicting the "cheap Chinese alternative" narrative that dominated 2024-2025 discourse. The shift suggests enterprises are increasingly planning to dual-source AI infrastructure across geographies and vendors, with performance and reliability becoming more important than raw cost arbitrage.

Anthropic voluntarily implements text watermarking and cryptographic signing under EU AI Act compliance

Anthropic has deployed embedded watermarks and signed provenance metadata on Claude text outputs, going beyond regulatory requirements by voluntarily signing the EU's Article 50(2) Code of Practice. This represents a concrete product implementation of AI output tracking that could become an industry standard as EU regulations spread globally.

Optiver shifts competitive moat from latency optimization to AI model quality as market efficiency increases

Proprietary trading firm Optiver has fundamentally reoriented its engineering strategy: latency optimization, once the primary competitive differentiator, has ceased to be a meaningful advantage as competitors have squeezed performance gains to near-parity. The firm now invests substantially more in building superior machine learning models than in lowering latencies, signaling a structural shift in how trading firms compete. According to Optiver's CTO Alex Itkin, the trading industry has evolved from automated mechanical workflows (late 1990s–2015) to quantitative, data-driven decision-making (2015–present), with information models and inference speed at network edges increasingly determining profitability. This mirrors broader AI infrastructure dynamics: as commodity latency becomes table stakes, proprietary model quality and inference efficiency become the real moat.

Figure's marketplace volume surges to $1.6B in Q1 2026 as partner-originated loans hit 78% of flow, margins expand sharply

Figure Technologies' marketplace volume reached $1.6B in Q1 2026, up from $8M in Q4 2024, with partner-originated loans now representing 78% of transaction flow. The shift from balance-sheet lender to capital-light marketplace is already materializing: ecosystem and technology fees have grown from 5% to 28% of revenue, and incremental EBITDA margins hit 91% in 2025. This economics improvement contradicts Street consensus on take-rate compression, suggesting Figure's unit economics are fundamentally improving even as first-lien mix (lower take-rate products) nearly doubled.

Chai Discovery closes four major pharma partnerships, signaling inflection from licensing to tools adoption in AI drug discovery

Chai Discovery announced partnerships with Eli Lilly, Novartis, and argenx after years in which AI-for-pharma companies defaulted to building their own drug pipelines. The inflection point: binding affinity models have become reliable enough that pharma teams now trust them for actual drug design workflows rather than just research validation, marking a shift from licensing deals to embedded tool adoption.

Researchers extract hidden reasoning traces from frontier AI APIs via cryptographic vulnerability, exposing secrets and distillation risk

Academics demonstrated a vulnerability in the APIs of Anthropic, OpenAI, and Google that allows extraction of encrypted reasoning blocks by replaying them across sessions and prompting weaker models to transcribe contents. Extracted token counts matched billed thinking tokens 1:1, confirming the attack's fidelity. A scan of ~7,000 public Claude/Codex sessions found 62 API keys, 33 email addresses, and 33 passwords leaked inside encrypted reasoning blocks. The finding highlights both a privacy exposure for users who share sessions and a potential distillation vector that labs believed they had mitigated—raising questions about the security model of reasoning APIs as they scale.

NVIDIA Nemotron 3.5 Lightning 30B MoE ships with 4× throughput gains, positioning open models as standard for agentic workloads

NVIDIA released Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3.6B active parameters, achieving 670 tokens/second serving speed and scoring 24 on Artificial Analysis' Intelligence Index with strong agentic performance (GDPval-AA v2 Elo 824). The model delivers up to 4× the output speed of similar-sized models and has rapidly distributed across Together AI, Ollama, Baseten, vLLM, and Perplexity API, positioning it as the open-model standard for always-on agent execution at scale.

Nathan Lambert's RLHF textbook reveals persistent gap in AI model writing capability despite coding and math advances

Nathan Lambert completed a 200–300 page technical textbook on reinforcement learning from human feedback in 2026, finding that despite massive progress in coding and math, current LLMs remain surprisingly weak at long-form non-fiction writing. Models struggle with organization, compound errors across sections, and fail to effectively compress and present established knowledge. This stagnation in technical writing is significant because it suggests fundamental limitations in how models utilize their knowledge on underspecified problems, raising questions about their ability to autonomously solve open-ended scientific problems without human guidance.

Ways to think about token pricing

Benedict Evans in Ben Evans published an essay titled "Ways to think about token pricing". Main takeaways:

• Token pricing today reflects a supply crunch, but all variables (supply, demand, marginal cost, ROI) remain highly uncertain over 3-5 year horizons, making confident forecasts unreliable.
• The critical unknowns determining frontier model pricing power are: (1) how many use cases justify frontier model costs vs. cheaper alternatives, (2) whether frontier capability improvements continue at rates that outpace efficiency/capacity gains, (3) whether competition remains fierce among 3-5 equivalent frontier labs or consolidates, and (4) how much value is captured by the model vs. tooling/process/data layers above it.
• Historical comparisons (fiber overbuild, mobile data, semiconductor fabrication, electrification) suggest frontier models may resemble low-margin commodity infrastructure where value concentrates upstream, though this is not predetermined.
• Inference has 40-50% gross margins today but excludes training costs (currently far larger than revenue); sustainability depends on unknown future training cost trajectories and actual ROI of emerging use cases.

NVIDIA's Risky Business

Stratechery published a strategic essay analyzing NVIDIA's competitive position and the AI infrastructure buildout through the lens of historical railroad financing. Main takeaways:

• Hyperscalers have shifted from self-funded CapEx to debt financing: Oracle, Meta, Alphabet, and Amazon raised $194B in debt through July 2026 (vs. $108B for all of 2025), with bond spreads rising and coverage falling to 2x from 5x in February
• Google's $85B equity raise and $10B Berkshire Hathaway investment signal expected compute demand far exceeds current supply; the company betting most cash on compute capacity may compound advantages in both capacity ownership and resale
• DeepMind leadership departures (Hassabis, Jeff Dean) and researcher exodus suggest Google's bureaucratic culture prevented competitive frontier model development, but this may actually benefit Google Cloud's infrastructure business under Thomas Kurian
• Anthropic represents 20%+ of Google's TPU shipments through 4Q27 and is the likely customer for Google's 'bridge deals' renting third-party capacity, making Google Cloud a major beneficiary even if DeepMind loses frontier status

Stealing Reasoning Traces from Proprietary LLM APIs

Panfilov, Schmotz, Shumailov, Beurer-Kellner, Schaeffer, Prabhu, Geiping, Andriushchenko in arXiv published a paper titled "Stealing Reasoning Traces from Proprietary LLM APIs". Main takeaways:

• Recovered 315,320 reconstructed reasoning blocks from 6,708 public agent trajectories, extracting 704 distinct privacy artifacts including 62 API keys, 33 passwords, and 24 access tokens that were hidden in reasoning traces
• Demonstrated reasoning traces can be replayed across different sessions and accounts using a single global encryption key, indicating potential cross-account security implications
• Decoded reasoning traces reveal "weirdness and casual misalignment" in frontier model thinking at scale, with implications for model monitoring and safety oversight
• Showed that prefilling one model's reasoning with another model's reasoning tokens influences visible outputs despite reasoning never being directly visible to users

With Taalas, AMD Can Bake AI Inference Directly Into Its Chippery

The Next Platform published an essay on AMD's acquisition of Taalas and its strategic implications for disaggregated AI inference architecture. Main takeaways:

• AMD acquired AI inference startup Taalas to build model-specific inference accelerators that hard-code model weights into ROM and pair with giant SRAM blocks for on-chip KV cache, enabling extremely low-latency inference at lower cost per token than Nvidia Blackwell B200 GPUs.
• Taalas HC1 chips store 8B-parameter models with HC2 supporting 20B parameters; tens of chips linked together can theoretically hold trillion-parameter models, with customization costs 100x cheaper than training new models.
• AMD's Taalas acquisition complements its Cerebras partnership for disaggregated inference, addressing GPU limitations on decode workloads that Nvidia demonstrated require specialized accelerators (like Grok) to achieve premium-tier agentic AI performance (1000+ tokens per second per user).
• This represents AMD's shift toward controlling its own inference acceleration stack as an alternative to expensive acquisition of Cerebras (currently $50.9B market cap at 300x annual revenue), positioning AMD to compete on inference economics as agentic AI workloads scale.

Claude Improves Riemann Zeta Function Lower Bound from 41.6% to 67.2%

Anthropic published a research report on Claude's mathematical discovery work. Main takeaways:

• An unreleased Claude research version improved the lower bound for Riemann zeta function zeros on the critical line from 41.6% to 67.2%, a significant advance on a decades-old mathematical problem.
• Claude achieved this by combining recent work from Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and Bombieri in a novel way, using quadratic form analysis without assuming the Riemann hypothesis itself.
• The discovery emerged from an agentic workflow: Claude coordinated ~60 subagents across two sessions, executing 2,400+ shell commands and writing hundreds of Python scripts with minimal human direction beyond encouragement.
• Claude produced a formally verifiable Lean proof and had the result validated by Anthropic's own mathematicians and external experts (Brian Conrey, Dan Goldston).
• The work consumed 31 million output tokens and demonstrates Claude's capability to extend mathematical research beyond its training, though the techniques are unlikely to resolve the full Riemann hypothesis.