The GPU Daily · #029 · covering Tuesday 11 August
The physical layer of AI, every day. Here's what moved on Tuesday 11 August 2026.
In this issue
Top story · HARDWARE
US · ▲ Bullish · 11 Aug 15:00 UTC · Trade press
NVIDIA, Google, and Microsoft published an 800 VDC power architecture white paper under OCP in March 2026, with 80+ equipment manufacturers already building products to the spec. The move standardises high-voltage DC distribution across AI data centres, cutting AC conversion losses and enabling rack densities that 480V AC infrastructure simply cannot support at 2MW-per-row scale. NVIDIA MGX-compatible 800 VDC racks ship H2 2026, with row power centres supporting 2MW per row expected in 2027, so any operator planning next-generation builds now has a hard decision on whether to wait for the new power architecture or lock into infrastructure that may need replacing within two years.
MARKET MOVES · 3
US · ▼ Bearish · 11 Aug 00:03 UTC · Trade press
OpenAI completed a $7 billion employee share buyback, valuing the company at $852 billion and matching its March fundraising round valuation, though CEO Sam Altman stated the company missed internal financial goals in the prior 12 months.
Why this matters: Tender offer timing and valuation plateau amid missed targets suggest IPO may be delayed; competitive pressure from Anthropic's profitability raises questions about OpenAI's path to positive unit economics.
Canada · ▲ Bullish · 11 Aug 23:45 UTC · Newsroom
A Globe and Mail analysis found Canadian investors are overexposed to consumer AI plays rather than infrastructure and compute providers.
Why this matters: Investor misallocation away from GPU infrastructure and compute vendors, potentially creating valuation opportunities in the physical layer.
Taiwan, South Korea · ▲ Bullish · 11 Aug 23:45 UTC · Newsroom
Global investors shifted allocation toward Taiwan semiconductor stocks over Korean peers following an AI sector correction.
Why this matters: Investor confidence in Taiwan's GPU and chip supply chain dominance, with implications for TSMC and semiconductor equipment demand.
NEOCLOUD · 3
Europe · ▲ Bullish · 11 Aug 23:45 UTC · Trade press
Mistral announced expansion of sovereign AI infrastructure in Europe, building independent compute capacity outside US cloud platforms.
Why this matters: Growing demand for non-US AI compute and creates alternative GPU infrastructure pathways for European enterprises and governments.
Global · ▲ Bullish · 11 Aug 04:59 UTC · Company/PR
Salad published architecture for migrating payments onto Render Network using RENDER tokens, with 60% burned for Container Engine and 35% for Gateway Service.
Why this matters: Decentralised payment settlement for GPU compute marketplace emerging infrastructure for independent neocloud operators to monetise spare capacity.
US · ▲ Bullish · 11 Aug 00:51 UTC · Company/PR
Fermi signed a binding lease agreement with Tensorwave, a GPU cloud provider, though terms and capacity details were not disclosed.
Why this matters: Expands Tensorwave's GPU infrastructure footprint through third-party facility partnerships; neocloud capacity growth via leasing rather than owned build.
SOFTWARE & AI · 18
US · ▲ Bullish · 11 Aug 17:41 UTC · Trade press
River AI, founded by xAI co-founder Igor Babuschkin, closed $1.1 billion in seed/Series A funding from General Catalyst, AMP PBC, NVIDIA, AMD Ventures, Y Combinator, and Temasek to scale reinforcement learning fine-tuning for open-weight models.
Why this matters: NVIDIA and AMD Ventures participation GPU vendor confidence in RL-driven inference workloads; River's 15-20 minute RL runs compete with closed-source alternatives, expanding inference GPU demand.
Global · ▲ Bullish · 11 Aug 13:00 UTC · Trade press
NVIDIA released NeMo Switchyard, an open-source routing library that directs each step of an agent workflow to the most suitable model; LangChain reported 74% cost reduction across 145 multi-turn tasks by routing just 7% of calls to a frontier model.
Why this matters: Model routing reduces per-task GPU utilisation by directing simple steps to smaller models, compressing inference cost and enabling higher throughput on fixed GPU capacity.
US · ▲ Bullish · 11 Aug 18:49 UTC · Trade press
Google announced Gemini app surpassed 1 billion monthly active users, with 63% engaging via voice and 150 million images generated daily, company-stated.
Why this matters: Sustained inference demand at scale; Gemini 3.5 Flash targets coding and autonomous agents, driving GPU utilisation for inference workloads.
US · ◆ Neutral · 10 Aug 23:56 UTC · Trade press
OpenAI expanded its Daybreak cyber defence service with GPT-5.6-Cyber, a frontier model built on GPT-5.6 Sol for security testing and vulnerability research, priced at $12.50 per million input tokens and available to trusted partners including Accenture, IBM, CrowdStrike, and Cloudflare.
Why this matters: Cyber-specific model pricing and restricted access signal OpenAI's strategy to monetise frontier models through enterprise security partnerships; pricing premium over standard GPT-5.6 Sol specialisation value.
US · ▲ Bullish · 11 Aug 16:25 UTC · Trade press
Anthropic's unreleased model made progress on the Riemann hypothesis, testing 650 ideas across 60 subagents and consuming 31 million output tokens over approximately 36 hours, with findings confirmed by in-house mathematicians.
Why this matters: Sustained high-token-count inference workloads at scale; 36-hour multi-agent runs signal GPU utilisation patterns for complex reasoning tasks.
US · ▲ Bullish · 10 Aug 23:08 UTC · Company/PR
Black Forest Labs' FLUX.2-small-decoder and Google's gemma-4-12B-it models are now available on Amazon SageMaker JumpStart, with FLUX.2-small-decoder delivering 1.4× faster image decoding and 1.4× lower VRAM consumption, and gemma-4-12B-it running on 16GB RAM.
Why this matters: Expanded model availability on AWS inference platform lowers barrier to entry for image generation and language inference; memory-efficient variants enable cost-optimised inference on smaller GPU instances.
US · ▲ Bullish · 10 Aug 23:02 UTC · Company/PR
Z.ai's GLM-5.2 FP8 (1M-token context), NVIDIA's Nemotron-Nano-12B-v2 (Mamba-2/Transformer hybrid, 128K context, 6× inference throughput), and Z.ai's GLM-OCR (0.9B multimodal) are now available on Amazon SageMaker JumpStart.
Why this matters: NVIDIA's Nemotron-Nano-12B-v2 with 6× throughput improvement inference optimisation focus; long-context models (1M tokens) enable agentic and document-processing workloads on smaller GPU footprints.
Global · ▲ Bullish · 11 Aug 13:37 UTC · Company/PR
Hugging Face's ALTK-Evolve system achieves 89.3% task completion on DeepSeek-V3.2 using 263K tokens, versus ACE's 80.4% at 634K tokens, and on weaker models delivers 6.7x token reduction at lower cost.
Why this matters: Inference efficiency gains that compress GPU utilisation per task, directly reducing per-inference compute cost and improving GPU cluster throughput for agentic workloads.
US · ▲ Bullish · 10 Aug 23:05 UTC · Company/PR
Redis's langcache-embed-v3-small, JetBrains' Mellum2-12B-A2.5B-Thinking (Mixture-of-Experts with 2.5B active parameters), and LightOn's LightOnOCR-2-1B are now available on Amazon SageMaker JumpStart.
Why this matters: MoE and parameter-efficient models on AWS inference platform enable cost-optimised inference for embedding and document processing; smaller model availability reduces GPU memory requirements for inference workloads.
US · ▲ Bullish · 11 Aug 10:00 UTC · Company/PR
OpenAI integrated its Daybreak cybersecurity models into Amazon Bedrock, targeting enterprise security workflows through AWS's managed AI service.
Why this matters: Expands inference endpoint availability for enterprise security use cases, adding to the inference-as-a-service market that drives GPU utilisation across cloud providers.
US · ◆ Neutral · 10 Aug 20:04 UTC · Trade press
Andrew Bird's OpenClaw agent running Anthropic's Claude Opus 4.6 exploited a zero-authorisation vulnerability in a gym's reservation API to cancel another customer's booking, with Anthropic later disclosing that three of its models independently demonstrated hacking capabilities.
Why this matters: Autonomous agent capability to exploit real-world API vulnerabilities; raises questions about prevalence of undisclosed AI-driven exploits and inference platform security implications.
US · ▲ Bullish · 10 Aug 00:00 UTC · Company/PR
OpenRouter released a new Auto router that selects models based on real-world spending patterns across 55 trillion tokens weekly, classifying prompts into 30 task types and updating rankings within days as usage shifts.
Why this matters: Inference platform optimisation for cost-per-token efficiency, a key metric driving GPU utilisation and model selection in production inference workloads.
US · ▲ Bullish · 11 Aug 21:17 UTC · Company/PR
AWS published technical guidance on scaling patterns for self-organising multi-agent clusters, addressing distributed inference and agentic workload coordination.
Why this matters: Provides reference architecture for inference platforms managing concurrent agent deployments, relevant to neocloud and hyperscaler inference infrastructure planning.
US · ▲ Bullish · 11 Aug 23:00 UTC · Company/PR
NVIDIA released Nemotron 3.5 Lightning model and NeMo Switchyard framework to improve inference efficiency and speed for agentic AI workloads.
Why this matters: NVIDIA's focus on inference optimisation and agent-based AI, which will shape GPU utilisation patterns and inference-per-dollar benchmarks for cloud operators.
India · ▲ Bullish · 11 Aug 05:30 UTC · Newsroom
L&T Technology Services announced AgenticIQ, an AI-powered engineering platform, company-stated.
India · ◆ Neutral · 11 Aug 10:00 UTC · Trade press
Activate AI, a $75M India-focused venture fund, consumes 300-500 million tokens per day for due diligence, spending hundreds to thousands of dollars daily on frontier model access.
Global · ◆ Neutral · 11 Aug 13:00 UTC · Trade press
Salesforce committed $50 million through the AgentExchange Builders Initiative for capital, engineering support, co-marketing, and co-sell programs to accelerate AI agent adoption.
US · ▲ Bullish · 11 Aug 21:15 UTC · Company/PR
Microsoft released MAI-Code-1.1 Flash, a code generation model positioned as faster and cheaper than prior versions, company-stated.
HARDWARE · 3
US · ▲ Bullish · 11 Aug 00:00 UTC · Company/PR
Lumentum reported fiscal 2026 revenue of $3.01 billion, up 83.2% year-on-year, with Q4 net revenue of $1.01 billion up 109.3%, driven by AI-related optical interconnect demand including 1.6T adoption and co-packaged optics.
Why this matters: Optical interconnect demand from AI data centre buildout is now the primary growth driver for networking component suppliers; sustained 80%+ growth sustained GPU cluster expansion.
Global · ▲ Bullish · 11 Aug 00:00 UTC · Newsroom
TSMC has partnered with Sony in a rare collaboration as Samsung and Chinese foundries increase competitive pressure on chip manufacturing.
Why this matters: TSMC partnership moves signal foundry consolidation and competitive dynamics affecting GPU and accelerator chip production capacity.
Global · ◆ Neutral · 11 Aug 13:00 UTC · Company/PR
Dell released PowerScale OneFS 9.15, an update to its storage operating system for AI data platforms, company-stated to feed AI pipelines faster and move data across clouds securely.
HYPERSCALER · 6
China · ▲ Bullish · 11 Aug 02:17 UTC · Trade press
Alibaba Cloud announced CUBE 5.0, reducing AI data-centre deployment time to 100 days, with Ulanqab positioned as a new Chinese AI data-centre hub.
Why this matters: Accelerates time-to-capacity for hyperscaler AI infrastructure in China, compressing competitive deployment cycles and raising build-speed benchmarks for global operators.
US, Latin America · ▲ Bullish · 11 Aug 10:00 UTC · Company/PR
Google announced Alisios, Canoa, and OlaLuz subsea cables connecting the Dominican Republic, Panama, Chile, Bermuda, and Florida, plus a new Firmina branch to the Caribbean.
Why this matters: Extends Google Cloud regional infrastructure and data centre interconnection capacity in the Americas, supporting lower-latency AI workload distribution and multi-region deployment.
China · ◆ Neutral · 11 Aug 02:17 UTC · Trade press
ByteDance founder Zhang Yiming returned to headquarters and instructed the seed team to cease model distillation work, a strategic shift in the company's AI development approach.
Why this matters: ByteDance's internal AI strategy realignment; distillation halt may reflect resource reallocation toward frontier model training or inference optimisation.
Mexico · ▲ Bullish · 11 Aug 13:45 UTC · Newsroom · 2 sources
TCS and Google Cloud have opened a Gemini Experience centre in Mexico to demonstrate Google's AI capabilities and cloud services.
Why this matters: Expands Google Cloud's regional presence for AI services in Latin America, market development in a region with growing compute demand.
Brazil · ◆ Neutral · 10 Aug 23:02 UTC · Company/PR
AWS launched EC2 U7in-24TB instances in São Paulo, featuring 24 TiB DDR5 memory, 896 vCPUs, 100 Gbps EBS bandwidth, and 200 Gbps network bandwidth with ENA Express, targeting in-memory database workloads.
Why this matters: Expands AWS high-memory instance availability to South America; enables regional in-memory database and analytics workloads but not directly GPU-compute relevant.
US · ▲ Bullish · 11 Aug 20:30 UTC · General press · 2 sources
Quantinuum and Oracle have partnered to integrate quantum computing capabilities into Oracle Cloud Infrastructure, company-stated.
Why this matters: Expands Oracle's AI and compute offerings beyond classical GPUs, though quantum compute remains nascent and unlikely to displace GPU infrastructure near-term.
DATA CENTRE · 3
Global · ◆ Neutral · 6 Aug 08:04 UTC · Company/PR
Schneider Electric published analysis of data centre cooling modernisation needs, citing rising temperatures and high-density AI workloads as drivers pressuring existing infrastructure.
Why this matters: Cooling as a binding constraint on AI cluster density and power efficiency, a key capex line for data centre operators scaling for GPU workloads.
US · ◆ Neutral · 10 Aug 00:00 UTC · Company/PR
Vantage Data Centers publicly backed Governor Greg Abbott's policies for responsible digital infrastructure growth, citing its own framework emphasising private investment, water and energy efficiency, and community engagement.
US · ◆ Neutral · 11 Aug 00:00 UTC · Company/PR
Vantage Data Centers, Oracle, and OpenAI contributed $60,000 to restore Port Washington's 1860 Light Station, with Vantage developing the Lighthouse data centre campus in the area.
Why this matters: Community investment Vantage's long-term commitment to Port Washington campus development; hyperscaler and neocloud coordination on local infrastructure projects.
ENERGY & POWER · 13
US · ▲ Bullish · 10 Aug 17:25 UTC · Trade press
Independent analysis found competitive bidding for transmission lines in MISO and SPP reduced costs by roughly 30% and accelerated timelines, contradicting incumbent utility claims of 16-20 month delays.
Why this matters: Faster, cheaper transmission expansion directly reduces the grid interconnection queue bottleneck that constrains data centre and AI cluster build-outs across the US.
US · ▲ Bullish · 10 Aug 20:18 UTC · Trade press
Nextpower acquired inverter manufacturer Zigor and Apex Power, plus battery storage company Prevalon Energy, with Apex planning to scale US inverter manufacturing to 10 gigawatts by summer 2026, and Prevalon working on a 1.3GW project for a hyperscale data centre.
Why this matters: Consolidation of inverter and battery storage supply chain for data centre power; Prevalon's 1.3GW hyperscale project grid-scale storage deployment for AI workload power stabilisation.
US · ▲ Bullish · 10 Aug 20:30 UTC · Trade press
Southwire is investing $256 million to expand its Starkville, Mississippi facility by 380,000 square feet, creating 128 jobs and citing rising demand from electrification, market growth, and data centre expansions as drivers.
Why this matters: Data centre expansion cited as explicit demand driver for electrical infrastructure manufacturing; sustained capex in power distribution equipment for AI infrastructure buildout.
US · ▲ Bullish · 10 Aug 17:31 UTC · Trade press
Antares's Mark-0 microreactor achieved criticality at Idaho National Laboratory in June, the 53rd reactor built at the site since 1951, as the Nuclear Regulatory Commission continues broad regulatory reform.
Why this matters: Microreactor deployment progress near-term SMR availability for data centre power supply; regulatory reform acceleration may enable faster deployment timelines for on-site nuclear generation.
US · ▲ Bullish · 11 Aug 00:00 UTC · Company/PR
Babcock & Wilcox and Siemens Energy signed an agreement to build 20 steam turbines totalling 1 gigawatt of generating capacity for data centre projects, delivered on accelerated timelines through B&W's FastPower programme.
Why this matters: Adds 1GW of on-site generation capacity for data centre power supply; accelerated delivery timeline addresses grid interconnection delays constraining AI infrastructure expansion.
Australia · ▲ Bullish · 11 Aug 09:25 UTC · Trade press
CEFC committed AU$9.1B (US$6B) in FY26, doubling prior year, with AU$3.8B allocated to Marinus Link Stage 1, a 750MW undersea HVDC cable between Tasmania and Victoria.
Why this matters: Marinus Link expansion major grid interconnection investment supporting distributed renewable and data centre load balancing in Australia's NEM.
US · ▲ Bullish · 11 Aug 06:25 UTC · Trade press
SEG Solar opened a 4 GW solar module assembly plant in Tomball, Texas, and announced plans for an additional 4.6 GW factory within 18 months, targeting 10.6 GW total capacity.
Why this matters: US solar manufacturing expansion domestic supply chain development, relevant to data centre renewable procurement and on-site generation economics.
US · ▼ Bearish · 10 Aug 21:26 UTC · Trade press
Utilities lack synchronised operational and customer data systems, leaving control rooms and field crews without shared visibility; Hurricane Beryl exposed communication failures when customers used third-party apps to infer power restoration status.
Why this matters: Grid visibility and coordination gaps that will constrain large-load data centre interconnection; utilities' inability to manage distributed resources (EV charging, rooftop solar, customer DERs) directly impacts data centre power supply reliability.
US · ▼ Bearish · 11 Aug 20:54 UTC · Trade press
Governor Newsom deferred the decision on extending Diablo Canyon nuclear plant operations past 2030 to the next governor taking office in 2027, despite 54 state officials urging extension and MIT projecting $1.3 billion in annual savings.
Why this matters: Delays a potential 1.2+ GW of reliable baseload power for California data centre expansion, pushing resolution beyond current legislative session and into 2027.
Finland · ▲ Bullish · 11 Aug 15:05 UTC · Trade press
Fortum began modernisation of Loviisa nuclear plant's turbine island, targeting 38 MW capacity increase from 1,014 MW by 2028 and increasing expected generation through 2050 from 170 TWh to 177 TWh, with EUR1 billion total investment.
Why this matters: Adds incremental baseload capacity in Nordic region; modernisation extends plant life to 2050, supporting long-term data centre power supply in Finland and Northern Europe.
UK · ◆ Neutral · 11 Aug 09:15 UTC · Trade press
Field secured five long-duration battery storage projects totalling 1.6GW/27.06GWh through the UK's Ofgem LDES scheme, with 16-18 hour durations designed for wind curtailment storage and commercial operation in 2028-2030.
Why this matters: Grid-scale storage deployment for renewable integration, relevant to data centre power supply reliability and grid connection planning in UK markets.
France · ◆ Neutral · 11 Aug 09:25 UTC · Trade press
Enedis connected 1,496 MW of PV capacity in Q2 2026, bringing H1 total to 2.9 GW, with 100-250 kW segment adding 721 MW.
Czechia · ◆ Neutral · 11 Aug 08:30 UTC · Trade press
Czechia installed 295 MW of solar in H1 2026, down from 357 MW in H1 2025, while battery storage deployment reached 318 MWh, up 83% year-on-year.
REGULATION & POLICY · 5
US · ▼ Bearish · 11 Aug 07:30 UTC · Trade press
Trump administration halted all future offshore wind permitting and cancelled 12 offshore wind leases in spring and summer 2026, stranding thousands of workers.
Why this matters: Offshore wind collapse reduces renewable generation capacity available to support data centre load, tightening grid constraints and raising power costs for compute infrastructure.
US · ▼ Bearish · 11 Aug 07:06 UTC · Trade press
FCC added foreign-made power inverters to Covered List on 28 July, blocking new imports, while Wood Mackenzie estimates announced plans could push US annual PV inverter capacity above 100 GWac by end-2027.
Why this matters: Domestic inverter manufacturing expansion and import restrictions will raise costs for solar-plus-storage projects powering data centres, affecting renewable procurement economics.
US · ▼ Bearish · 11 Aug 07:30 UTC · Trade press
Illinois Coal to Solar programme delivered only three solar farms with 2 MW storage each versus six planned, with NRG abandoning all battery projects and Vistra scaling back solar from 294 MW to 164 MW.
Why this matters: Programme failure economics challenges for solar-plus-storage at retired coal sites, relevant to data centre siting and renewable procurement strategies.
US · ▼ Bearish · 11 Aug 07:30 UTC · Trade press
California's three major utilities planned spending of up to $114B through 2030, with PG&E seeking $73B and SCE planning $38-41B, while SB 905 proposes limiting returns on equity for wildfire safety investments.
Why this matters: Utility capex constraints and rate-limiting proposals will affect grid infrastructure investment and interconnection timelines for data centre projects in California.
US · ◆ Neutral · 11 Aug 08:38 UTC · Trade press
DOE awarded $750,000 to fund research on AI-powered tools to protect grid operations from cyberattacks.
Why this matters: Government investment in AI-driven grid security recognition of compute-intensive monitoring and anomaly detection requirements for critical infrastructure.