Why AI productivity gains aren't translating to firm profits

Microsoft and Nvidia unveil Windows PCs with custom Nvidia SoCs; SoftBank commits €75B to European AI infrastructure

Microsoft and Nvidia are introducing the first Windows PCs built on Nvidia's custom processors, including devices from Surface and Dell, launching at Computex and Build 2026. This represents a significant competitive move against Apple's chip dominance and creates a new hardware category for on-device AI compute, particularly relevant for power users running LLMs locally. Separately, SoftBank is leading a major European AI infrastructure buildout in Hauts-de-France with up to €75B in commitments, including €45B dedicated to 3.1GW capacity by 2031, demonstrating how capital is flowing to secure compute capacity outside the US and signaling intensifying competition for data center resources and power grids globally.

DeepSeek V4 and Huawei Ascend demonstrate CUDA alternative at scale with 65% model flops utilization

DeepSeek V4's collaboration with Huawei Ascend validates that alternatives to Nvidia's proprietary software ecosystem can achieve production-grade reliability and performance, with AIGCode reporting 65% Model Flops Utilization (nearly double industry average) on Ascend chips during MoE model pre-training. This matters because CUDA's moat has been central to Nvidia's competitive advantage; if Huawei's open-sourced CANN framework can sustain organic developer contribution and ecosystem growth, it could materially reshape China's independent compute capacity. CANN has transitioned from "infancy" to "youth" phase with 4M+ developer community members and 70+ major AI models supported following its December 2025 open-sourcing. The framework's shift from vendor-dependent support (previously 4-month wait times for bug fixes) to independent developer problem-solving is evidenced by concrete signals including a major Chinese bank integrating AI into core risk management and contributing 34 optimizations to vLLM-Ascend.

Anthropic hits $4.8B quarterly revenue with 80x growth; AI sector reaches $25B quarterly revenue with only one bubble indicator in red

Anthropic achieved $4.8B quarterly revenue (annualized $19.2B) with 80x year-over-year growth, shocking even CEO Dario Amodei who stated the company "tried to plan very well for a world of 10x growth per year. And yet we saw 80x." This explosive acceleration forced rapid infrastructure scaling and compute constraints. Across the broader AI sector, quarterly revenues have reached $25B (up from $13B in September 2025), with OpenAI growing from $1.7B to $6B annually. According to Azeem Azhar's analysis of five empirical bubble indicators derived from 300 years of investment booms and busts, only industry strain remains in the red zone. Revenue momentum is now in strong green territory with doubling time at 0.73 years, while economic strain has entered amber (AI capex now at 1% of US GDP, matching late-1990s telecom peak). Valuation multiples remain cool at 33x NASDAQ P/E versus 72x at dot-com peak, and if revenue growth holds on trend, industry strain should exit red by end-2026.

US AI economy growing 2,600% annually but invisible in GDP statistics, creating fiscal policy blind spot

University of Virginia, Anthropic, and Bank of Canada researchers estimate nominal AI GDP at $250B in 2025 with quality-adjusted growth at 2,600% per year, yet this remains largely unmeasured in conventional statistics because per-unit AI prices fall as fast as capabilities improve. This matters because policymakers relying on standard GDP data will systematically underestimate labor displacement risks and be unprepared for tax-base shocks requiring major fiscal reforms.

Biohub releases ESMFold2 protein model outperforming DeepMind's AlphaFold 3

The Chan Zuckerberg-backed Biohub released ESMFold2 alongside ESMC (trained on 2.8B protein sequences) and ESM Atlas (1.1B predicted structures), demonstrating scaling laws where inference-time sampling improves antibody-antigen pass rates from 49% to 65%, with lab-validated protein binder designs achieving 36-88% hit rates for cancer immunology targets. This represents concrete progress in AI-driven drug discovery and demonstrates positive-sum competition between major AI developers on real-world applications.

Why AI isn't showing up on your bottom line

Exponential View published an essay on why individual AI productivity gains are not translating to firm-level ROI despite widespread deployment. Main takeaways:

• Only 27% of executives report AI has met ROI expectations; individual worker productivity is up but organizational gains remain elusive ("one plus one plus one equals one-and-a-half")
• AI adoption mirrors electrification's three-stage pattern: Stage 1 (lightbulb—individual task speedup), Stage 2 (group drive—workflow automation on existing processes), Stage 3 (unit drive—firm-level reorganization around decision speed)
• Most current AI deployments are stuck in Stage 1-2; progression to Stage 3 requires rebuilding decision-making layers and managerial oversight, not just faster workflows
• "Congestion" occurs when individual/team output accelerates but decision pipelines remain unchanged; adding more AI output to blocked processes makes things worse
• Stage 3 requires AI to make decisions autonomously (e.g., feature requests flowing directly from customer signal to code without human intermediaries), fundamentally restructuring how firms operate

Can Huawei Take On Nvidia's CUDA?

ChinaTalk published an essay analyzing Huawei's three-pronged strategy to challenge Nvidia's software moat in AI hardware. Main takeaways:

• Huawei is deploying CANN (Compute Architecture for Neural Networks) as a CUDA alternative, paired with MindSpore framework, but developers report serious usability issues and instability compared to Nvidia's mature ecosystem
• Huawei is embedding engineering teams directly at customer sites (Baidu, iFlytek, Tencent) to assist code migration from CUDA to CANN, mirroring Nvidia's early CUDA adoption strategy from the 2000s
• Huawei is investing in PyTorch backend support (torch_npu adapters) and ONNX standards to reduce switching friction, but lacks the robust developer community and troubleshooting infrastructure that makes CUDA's ecosystem defensible
• Developer forums for CANN show minimal engagement compared to Nvidia's active CUDA communities, with one Zhihu commenter estimating CANN won't reach true maturity until 2027
• Model adaptation for Ascend requires deep optimization by Huawei engineers and progresses slowly, whereas Nvidia's out-of-box performance requires minimal vendor support

Tracking Trillions: The Assumptions Shaping Scale of the AI Build-Out

Goldman Sachs published a report analyzing the structural assumptions underlying AI infrastructure investment estimates. Main takeaways:

• AI infrastructure capex scale is most sensitive to three factors: silicon useful life assumptions, data center cost/complexity, and build-out composition/timing—not training vs. inference mix or memory pricing volatility
• Data centers designed less than two years ago may be obsolete for next-gen chips due to power/cooling demands, creating risk that historically durable assets become liabilities if architectural requirements shift materially within years of commissioning
• Shift from GPUs to ASICs depends critically on demand elasticity: if demand is inelastic, cheaper silicon reduces total capex; if elastic, lower costs unlock more usage and reshape value distribution rather than total spend
• Elongation (deployment delays from power queues, permitting, labor shortages, equipment lead times) is primarily a timing/volatility problem in base case, but becomes feedback loop if bottlenecks shift narrative from supply-side to demand-side questions about revenue justification
• Current ~$7.6 trillion infrastructure estimates are highly conditional on assumptions that may shift; innovation (like DeepSeek-style efficiency breakthroughs) could materially alter investment landscape

Where is AI in GDP Statistics?

Peterson Institute for International Economics in Peterson Institute for International Economics (PIIE) published a report titled "Where is AI in GDP Statistics?". Main takeaways:

• Quality-adjusted AI production in the US grew over 2,000% annually in 2024-2025, driven by data-center capacity expansion, hardware efficiency gains, and algorithmic progress.
• Preliminary nominal AI GDP estimates reach ~$250B in 2025, with quality-adjusted real growth at ~2,600% annually—largely invisible in conventional GDP statistics.
• National statistical accounts were not designed to track AI-sector activity; agencies should develop AI-focused satellite accounts now to prevent measurement gaps from becoming policy gaps.