Today's News
Azeem Azhar, Exponential View
Token consumption has exploded 17,000x over four years, forcing enterprises to exhaust annual AI budgets within months. Uber and ServiceNow both depleted their entire 2026 token budgets within four months, while average monthly AI spend at large enterprises jumped 36% to $85,521 between 2024 and 2025. The FinOps Foundation reports that 71% of companies experienced AI/ML cost overruns in 2025, with generative AI becoming the least predictable cloud category (48% of respondents). Cost management has become the top concern for 52% of CFOs, yet only 51% of organizations can confidently evaluate AI ROI—a critical gap given that 88% of software companies now use generative AI in production workflows. The core problem: OpenAI invoices show only aggregate token counts, not which customers, features, or workflows generated the spend, meaning only 32% of companies can allocate even half their OpenAI bill. Without granular cost attribution and real-time visibility, the shift from predictable SaaS economics to highly variable consumption-based spending is creating unsustainable financial planning challenges across the enterprise.
Techmeme
Analysis from The Information shows that 34 leading AI startups are generating approximately $80 billion in annualized revenue, up 112% in six months. Anthropic and OpenAI alone capture 89% of this revenue, providing concrete evidence of extreme market concentration and the durable economic moat of frontier labs. This metric is significant for understanding competitive dynamics and the structural advantages of scale in AI infrastructure and model development.
Techmeme
Power prices on the largest electric grid in the US—the 13-state PJM Interconnection—jumped 76% in Q1 2026 to $136.53 per megawatt-hour, compared to $77.78 in Q1 2025, driven by rampant demand from AI data centers. This acute infrastructure constraint directly impacts operating costs for data center operators and signals a critical bottleneck that could affect AI model training economics and deployment strategies across the industry.
Jack Clark from Import AI
Tilde Research identified that Muon causes permanent neuron death in MLP layers during training warmup, then developed Aurora as a leverage-aware alternative optimizer. Aurora achieved 2.26 final loss versus Muon's 2.31 on 1.1B-parameter models trained on ~100B tokens, demonstrating 100x data efficiency on open-source internet data. Independent validation from Pleias confirmed Aurora's superiority over both Muon and AdamW at 600M scale. Aurora is a drop-in replacement with only 6% overhead and has achieved state-of-the-art on the modded-nanoGPT optimization track, with code open-sourced on GitHub.
ChinAI Newsletter
The China Integrated Circuit Industry Investment Fund has made its first public backing of an LLM player, valuing DeepSeek at $50 billion or higher. The move signals strategic commitment to domestic AI stack autonomy, with DeepSeek V4 optimized for Huawei Ascend chips framed as a breakthrough in chip-model partnership. This reflects a broader geopolitical bet on breaking foreign technology dependencies rather than near-term commercial returns.
Latent.Space
Yaroslav Azhnyuk, founder of Petcube, has transitioned from pet cameras to founding one of the world's leading on-drone AI companies, now supplying autonomous modules, thermal cameras, and strike/interceptor drones to 200+ Ukrainian manufacturers and the armed forces. This demonstrates how commercial AI talent is being directly applied to autonomous weapons systems in active warfare, with implications for how AI infrastructure gets deployed in conflict.
Worth Reading
CNBC
I read the full report. Main takeaways:
- U.S. Commerce Department approved ~10 Chinese firms (Alibaba, Tencent, ByteDance, JD.com) to buy Nvidia H200 chips, each permitted up to 75,000 units, but zero deliveries have occurred despite approvals
- Chinese government is actively blocking or vetoing purchases internally despite U.S. approval, citing strategic concern that imports would weaken domestic AI chip development efforts
- Trump negotiated 25% revenue-sharing arrangement requiring chips pass through U.S. territory before shipment to China; Beijing views this with suspicion over potential tampering
- Nvidia's China market share has effectively fallen to zero under export controls; company previously commanded 95% of China's advanced chip market and China represented 13% of Nvidia revenue
- Beijing's supply chain security regulations now require government-wide effort to eliminate foreign dependencies in critical tech infrastructure, directly blocking H200 purchases
Read moreChinAI Newsletter (translation of Shuzili chang)
I read the full essay. Main takeaways:
- DeepSeek's valuation jumped from $10B to $51.5B in under 20 days, driven by strategic scarcity: it's building the first cutting-edge LLM ecosystem independent of U.S. semiconductor technology via Huawei Ascend chip integration.
- The China Integrated Circuit Industry Investment Fund (Big Fund) led the surge by investing in a software company for the first time, signaling state-level validation of DeepSeek as a strategic asset for technological sovereignty.
- DeepSeek's valuation premium reflects a shift from pure commercial metrics (like OpenAI's 35x ARR multiple or Anthropic's enterprise growth) to a "strategic scarcity" thesis: it's positioned as China's "Android" for AI, analogous to Huawei's role in semiconductors.
- V4's hardware validation checklist includes both Huawei Ascend and NVIDIA GPUs, but engineering costs of operator-by-operator migration to CANN delayed V4 release—indicating real technical friction in domestic chip substitution, not just marketing.
Read moreTilde Research
I read the full technical report. Main takeaways:
- Muon optimizer causes neuron death in MLP layers on tall matrices due to row-norm anisotropy; leverage scores explain the self-reinforcing feedback loop that prevents dead neurons from recovering
- Aurora optimizer solves the tension between row-norm uniformity and orthogonality by formulating steepest descent under joint constraints, achieving 100x data efficiency at 1.1B scale
- Aurora outperforms current SoTA on modded-nanoGPT speedrun (beats NorMuon) with only 6% computational overhead and is a drop-in replacement for Muon
- Row normalization in NorMuon necessarily introduces precision defects for tall matrices (mathematical proof provided), but Aurora recovers orthogonality precision while preventing neuron death
- Open-source implementation released; results suggest Aurora's gains scale with MLP width, making it particularly effective for networks with large MLP expansion factors
Read more