Why nuclear stagnation is America's growth problem

ECTC 2026: Intel EMIB-T, Marvell Custom HBM, and Direct-to-Silicon Cooling Reshape AI Accelerator Packaging

Advanced packaging has become the primary scaling vector for AI accelerators as transistor density gains slow. At ECTC 2026, Intel validated EMIB-T at 36µm bump pitch on 2× reticle silicon—a 65% increase in bump density over Granite Rapids—with 4.5× reticle certification targeted by end of 2026. Intel's roadmap includes sub-25µm bump pitch and embedded voltage regulators, positioning EMIB-T as a credible alternative to TSMC's CoWoS for large packages, though Intel remains behind on integrated voltage regulators and active silicon interconnect. Marvell demonstrated custom HBM4E that reduces host ASIC footprint dedicated to HBM PHYs by ~60% while shortening interposer channel length from 6.5mm to 1.5mm, enabling higher bandwidth on organic RDL interposers at lower cost—a strategy Nvidia is adopting for Feynman. TSMC and Microsoft demonstrated direct-to-silicon microfluidic cooling dissipating 5+ kW on test vehicles and actual Nvidia GH200 GPUs respectively, with TSMC's micropillar approach reaching 5.3 kW at 8 LPM and Microsoft reporting 50% reduction in package thermal resistance across real workloads. These disclosures signal that multi-kilowatt packages are overwhelming conventional cooling architectures, making advanced thermal solutions critical infrastructure constraints for next-generation accelerators.

Smart Model Routing Emerges as Competitive Battleground for AI Cost Optimization

Factory AI, Not Diamond, OpenRouter, and a dozen other vendors are launching intelligent routing solutions that automatically select cheaper models for simpler tasks, with Factory AI reporting enterprise demand "off the charts" and claiming 20-25% cost savings. Factory Router maintains frontier performance on 99% of tasks while cutting token spend by 20-25% by routing simple work (documentation, small bug fixes, mechanical refactors) to efficient models and reserving frontier models for tasks requiring deep reasoning. This addresses a critical pain point as token prices vary 10-20x between frontier and open models, with Factory's CEO estimating open models are sufficient for ~60% of coding work. The Pareto frontier analysis shows cost drops sharply while performance barely moves near the top of the curve, enabling aggressive routing without sacrificing quality on the work that matters.

Meta Contracted 5GW+ Compute in H1 2026, Pursuing Four Distinct Monetization Paths Beyond Internal AI

Meta has contracted over 5GW of capacity across cloud and colo deals in the first half of 2026, with two massive datacenter campuses representing 2.5GW under construction, contradicting market narratives about overcapacity. Rather than becoming a traditional neocloud, Meta is pursuing four high-value use cases: frontier model training via Meta Superintelligence Labs (MSL), scaling recommendation systems by >10x complexity, a Bedrock-style partnership with Anthropic for private Claude instances, and SpaceX-model on-demand compute deals at $50B/GW annual revenue. This optionality—particularly the ability to rapidly repurpose compute between research and revenue-generating services—makes aggressive capacity contracting economically rational even if MSL underperforms, with bulk capacity additions now sourced from third-party providers like Coreweave and Nebius. Meta's GEM ads foundation model demonstrated 4x better compute efficiency gains and 23x effective training FLOPs improvement, enabling profitable absorption of >10x AdRec compute increases while sustaining ROAS growth. Q1 2026 showed 19% YoY impression growth and 12% YoY average price-per-ad growth, with Instagram conversion rates up 5% after doubling GEM training GPUs, suggesting RecSys scaling remains a viable high-margin alternative to frontier model development.

Agent Adoption Reaches 95% Among AI Engineers, But Control Mechanisms Remain Primitive

Amplify's 2026 survey shows 95% of AI engineers now use agents—roughly double last year's share—with 89% deploying agents that take direct actions in systems, up from 52% the previous year. Agent capabilities have shifted dramatically from read/summarize/draft tasks to direct system manipulation, yet control mechanisms remain comparatively primitive: human approvals and permissions dominate safeguards, followed by scattered task decomposition, retrieval, memory, and sandboxing techniques. "Nobody has settled the control layer for agents," according to the survey, revealing a critical gap between capability deployment and governance maturity. Vercel's launch of eve, a prescriptive agent framework, signals the industry's pivot toward agent-native infrastructure, bundling primitives for context, tools, resumability, and long-running work—areas Vercel identifies as fundamentally different from web application infrastructure. Cost remains a concern, with 40% of respondents saying AI costs regularly limit how ambitiously they use AI, while 59% fear that today's AI-generated code is creating long-term liabilities.

Microsoft Launches $2.5B Frontier Company to Embed 6,000 Engineers Inside Enterprise Customers

Microsoft announced "The Microsoft Frontier Company," a $2.5 billion forward-deployed engineering initiative that will embed 6,000 engineers inside customers to build and run AI systems on-site. Led by Rodrigo Kede Lima, the initiative positions Microsoft as the infrastructure and services layer for enterprise AI deployment, enabling organizations to build proprietary AI capabilities while compounding human and token capital. This move directly competes with OpenAI's $4B+ Deployment Company (backed by TPG and others), Anthropic's $1.5B venture with Goldman Sachs and Blackstone, and Amazon's $1B forward-deployed engineering commitment announced days earlier. Microsoft is pitching privacy, trust, and model choice as differentiators—customers can run whichever AI model fits the job from OpenAI, Anthropic, Microsoft, or open-source providers—though the risk remains that working with Microsoft's engineers naturally embeds customers into Microsoft's cloud platform and technologies, making competitive switching difficult.

America's Next 250

Not Boring published an essay by Scott Nolan on America's next 250 years of progress, with focus on energy as the civilizational bottleneck. Main takeaways:

• U.S. nuclear generation stagnated for 23 years (2001-2023): 768M to 775M MWh, <1% growth despite superior safety/emissions vs. wind and solar
• Real wage growth per capita halted post-1970 oil shock; America fell off the "Henry Adams Curve" of 7% annual energy growth and never recovered
• Energy abundance is prerequisite for prosperity; nuclear power is identified as the critical infrastructure bottleneck to solve for next 250 years of U.S. growth
• Nolan frames nuclear as clean, safe, energy-dense, reliable, and scalable—positioning it as essential infrastructure alongside historical technological breakthroughs (electricity, internet, genomics)

Nvidia – The Inference Kingdom Expands

SemiAnalysis published a detailed technical report on Nvidia's GTC 2026 inference announcements and architecture strategy. Main takeaways:

• Nvidia acquired Groq's IP and team for $20B, integrating Groq's LPU (Latency Processing Unit) into Nvidia's Vera inference stack as LP30/LP35/LP40 chips to accelerate decode-phase latency in high-interactivity scenarios
• LP30 uses Samsung SF4 node with 500MB on-chip SRAM (vs LPU1's 230MB), enabling production without consuming constrained TSMC N3 or HBM allocations; LP40 will move to TSMC N3P with CoWoS-R and hybrid-bonded DRAM from SK Hynix
• Attention FFN Disaggregation (AFD) technique maps stateless FFN operations to LPUs (deterministic, low-latency SRAM) while keeping stateful attention on GPUs (dynamic KV cache), improving decode throughput and latency in MoE models
• Rubin generation scales NVLink bandwidth to 14.4 Tbit/s per GPU with bi-directional SerDes; Feynman NVL1152 multi-rack system introduces CPO (co-packaged optics) for scale-up networking to overcome copper's 2-meter range limit
• CPO reduces transceiver power by 73% (16-17W DSP transceiver vs 4-5W optical engine+laser) and enables larger scale-up domains beyond single-rack constraints

Stop Saying Half of 2026 US Datacenter Capacity Is Canceled

SemiAnalysis published a detailed report debunking widely-circulated claims that 50% of 2026 US datacenter capacity is delayed or canceled. Main takeaways:

• SemiAnalysis' datacenter tracking model shows only ~1% movement in YE2026 NA hyperscaler self-build forecasts and <5% in colocation over 6 months, contradicting Bloomberg/Sightline's 50% delay narrative
• The 50% figure relies on flawed methodology: Sightline's 12GW 2026 denominator undercounts actual under-construction capacity by multiples; top two hyperscalers alone exceed Sightline's 5GW under-construction estimate
• Most flagged delays are speculative early-stage projects without financing, interconnection studies, or equipment orders; these belong in 2028+ timelines, not 2026
• Real delays do occur (STACK Infrastructure/Oracle now delayed to 2029 due to gas pipeline constraints; Nebius New Jersey slipped from 4-month to 6+ month timeline) but are exceptions, not the rule
• Vibe-coded AI forecasts (Claude-generated models) systematically misinterpret press releases and construction timelines, creating false delay narratives that haven't moved SemiAnalysis' equipment supplier forecasts

AIEWF Daily Dispatch: The great loops debate and the state of AI engineering

Latent Space published a report on the AI Engineer World's Fair, capturing the central debate about autonomous software factories and agentic loops, alongside concrete survey data on AI engineering practices. Main takeaways:

• 95% of AI engineers now use agents (up from ~50% prior year); 89% report agents can write data, up from 52%, indicating shift from read/summarize to action-taking systems
• Control layer remains primitive: human approvals and permissions dominate safeguards; "nobody has settled the control layer for agents" per Amplify survey
• 76% of engineers report AI costs constrain ambition (40% regularly, 36% sometimes); token usage is now second-most monitored production metric after quality
• Anthropic's Claude Tag shows internal adoption pattern: 65% of product team code created by internal Tag version; enables async, multiplayer, proactive delegation vs. single-user chat
• Core tension: 59% fear AI-generated code creates long-term liabilities despite enabling more software production; debate reflects discipline lagging behind hype on autonomous loops