The GPU Daily · #028 · covering Monday 10 August
The physical layer of AI, every day. Here's what moved on Monday 10 August 2026.
In this issue
Top story · HARDWARE
South Korea · ▲ Bullish · 10 Aug 00:30 UTC · Newsroom
SK Hynix is expanding HBM production capacity at its South Korean facilities, a direct response to the memory bottleneck that has constrained H100, H200, and B200 cluster deployments at scale. HBM supply, not GPU silicon, has been the binding constraint on large cluster builds for the better part of two years, and any meaningful capacity addition shifts the calculus for operators planning 2025 and 2026 procurement. Watch for Samsung and Micron to respond with their own capacity announcements within the quarter, and for spot HBM pricing to soften if Hynix's ramp outpaces near-term absorption.
MARKET MOVES · 2
US · ▲ Bullish · 10 Aug 20:01 UTC · Trade press
NVIDIA signed memorandums of understanding with six major asset managers and financial firms to establish dedicated compute financing platforms targeting over $500 billion in third-party capital for AI infrastructure.
Why this matters: Formalises NVIDIA's positioning of GPU compute as an investable asset class; unlocks institutional capital for neocloud and hyperscaler buildout, accelerating capacity deployment.
US · ▲ Bullish · 10 Aug 15:00 UTC · Newsroom · 2 sources
Intel announced a $15B secondary offering as its stock rebounds amid restructuring efforts, confidence in its foundry and manufacturing strategy.
Why this matters: Raises capital for foundry expansion and process node roadmap execution, directly affecting Intel's ability to compete in server CPU and custom silicon markets.
NEOCLOUD · 1
Global · ▲ Bullish · 10 Aug 08:00 UTC · Company/PR
Civo offers NVIDIA B200, H200, H100, A100, and L40S GPUs on its platform, with B200 becoming available by August 2025, company-stated.
Why this matters: Expands GPU availability on independent cloud platform, increasing competition with hyperscalers on inference and fine-tuning workloads.
SOFTWARE & AI · 16
US · ▲ Bullish · 10 Aug 16:22 UTC · Trade press
Meta released Muse Glimmer, a 30-billion-parameter open-weight model under Apache 2.0 license, designed to run locally on single consumer GPUs (RTX 3090/4090/5090) for multi-step agentic tasks without internet.
Why this matters: Open-weight agent model targeting consumer GPU inference compresses the cost of local agentic compute and Meta's strategy to distribute personal superintelligence to individuals rather than centralised cloud.
US · ◆ Neutral · 10 Aug 18:31 UTC · Trade press
Brex released CrabTrap, an open-source HTTP proxy that monitors outbound network traffic from AI agents using LLM-based policy enforcement, routing 98% of routine actions through static rules to avoid latency.
Why this matters: Enterprise adoption of agentic systems at scale and the infrastructure challenge of securing code-execution agents, likely to influence how cloud providers architect agent sandboxing.
Global · ▲ Bullish · 10 Aug 02:08 UTC · Trade press
Meta released Muse Glimmer, a 30B parameter open source multimodal model with 2B vision encoder and 28B text decoder, available on Hugging Face with day-0 support in transformers, vLLM, and llama.cpp.
Why this matters: Open-weight multimodal model with native NVIDIA, AMD, and Intel GPU support lowers inference deployment friction and expands accessible inference endpoints.
US · ◆ Neutral · 10 Aug 20:00 UTC · Trade press
AWS launched Continuum, a code vulnerability platform integrating OpenAI Codex and Anthropic Claude, with AWS absorbing token costs and charging customers a single price regardless of underlying model selection.
Why this matters: Hyperscaler willingness to absorb frontier model inference costs as a competitive feature; the economics of inference workloads are shifting toward bundled pricing and volume absorption.
Global · ▲ Bullish · 10 Aug 10:05 UTC · Company/PR
Hugging Face and Multiverse Computing published a paper reducing knowledge distillation VRAM from 250GB to 128GB peak, and fused chunked KL loss cut memory from 85.2GB to 5.45GB at 32K token context, enabling Llama 3.1 8B distillation on one GPU node instead of four.
Why this matters: Reduces GPU requirements for model distillation at scale, lowering inference cluster capex and enabling smaller operators to compete on cost-per-inference with hyperscalers.
US · ◆ Neutral · 10 Aug 14:30 UTC · Trade press
Pete Johnson, Field CTO at MongoDB, argues that agentic memory systems backed by semantic search will replace token-maxxing as the core architectural pattern for AI agents in 2026, based on 100+ customer conversations.
Why this matters: Shift in enterprise agentic architecture from large context windows to persistent queryable memory, reducing inference cost per query and potentially lowering GPU utilisation per agent.
China · ◆ Neutral · 10 Aug 08:15 UTC · Trade press
China's large language models led global token usage for 15 consecutive weeks, with DeepSeek-V4-Flash taking the top spot, company-stated.
Why this matters: Sustained demand for Chinese inference endpoints and training clusters, domestic AI infrastructure investment is scaling faster than Western alternatives in some segments.
China · ▲ Bullish · 7 Aug 13:40 UTC · Company/PR
MiniMax H3 ranked second on Artificial Analysis video-generation leaderboard at 1,238 Elo versus Veo 3.1 at 1,093 Elo, with H3 priced at $7.80 per minute versus Veo 3.1 at $24.00 per minute.
Why this matters: Open-weight video model with superior price-to-performance ratio and native audio support compresses inference-per-dollar economics for video generation workloads.
US · ▲ Bullish · 10 Aug 05:30 UTC · Newsroom
Anthropic has turned on auto mode for Claude Code by default, expanding the inference platform's autonomous code generation capabilities.
Why this matters: Increases inference load per user session as Claude Code auto mode runs more compute-intensive operations without explicit user prompts, raising per-user GPU demand.
US · ▲ Bullish · 10 Aug 18:34 UTC · Company/PR
Cloudflare held Agents Week (August 3-7, 2025) introducing Agent Cloud with execution primitives, development lifecycle tools, zero-trust security, and an Agentic Internet framework.
Why this matters: Positions Cloudflare's global network as infrastructure for agent deployment, competing with hyperscaler-hosted agent platforms and demand for edge-distributed agentic compute.
Global · ▼ Bearish · 5 Aug 10:42 UTC · Company/PR
MiniMax H3 offers a globally available hosted API with no territorial restrictions, but self-hosted weights exclude the EU, UK, South Korea, and the US, with self-hosted deployment requiring prior authorisation above $20M yearly revenue.
Why this matters: Ongoing copyright litigation risk for open-weight model deployment; creates a two-tier licensing model that may influence how other foundation model providers structure commercial access in regulated markets.
US · ◆ Neutral · 10 Aug 16:20 UTC · Trade press
Meta released Muse Glimmer, a 30B open-weight model trained across 100+ languages, designed for on-device agentic tasks with text and image support, while keeping more powerful Muse Spark closed-weight.
Global · ▲ Bullish · 6 Aug 00:00 UTC · Company/PR
Nebius released engineering techniques for training speculative decoder draft heads, reducing peak memory from 501 GiB to 0.97 GiB in Llama 3.1 8B inference workloads via Streaming Cross Entropy and Block-sparse FlashAttention.
Why this matters: Lowers the memory and compute cost of speculative decoding, a key inference optimisation technique; enables smaller clusters to run larger batch sizes or longer sequences, compressing inference-per-token costs.
Global · ◆ Neutral · 7 Aug 12:44 UTC · Company/PR
Atlas Cloud paired Seedance 2.0 Mini (draft at $3.20 per million tokens) with Seedance 2.5 (final render at $9.70 per million tokens) to reduce compute costs by approximately 70% across multi-asset campaigns.
Why this matters: Two-tier video generation workflow cost-optimization patterns for inference platforms and tiered pricing models for compute-intensive workloads.
US · ▲ Bullish · 10 Aug 04:30 UTC
OpenAI acquired NextSlide to integrate AI slide-deck creation into ChatGPT, company-stated.
Why this matters: OpenAI's focus on consumer and enterprise productivity tools; does not directly impact GPU infrastructure demand or compute procurement.
China · ◆ Neutral · 7 Aug 00:00 UTC · Company/PR
Alibaba released Wan3.0, a video generation model with doubled output duration and expanded input modality support.
HARDWARE · 9
Taiwan · ▲ Bullish · 10 Aug 10:15 UTC · Trade press
TSMC reported July revenue of $14.5 billion, a 44.7% year-on-year jump, and guided to 2026 growth exceeding 40%, driven by AI accelerator and advanced node demand.
Why this matters: Confirms sustained demand for leading-edge foundry capacity for AI chips; 40%+ growth guidance multi-year visibility into GPU and custom accelerator orders from hyperscalers and neoclouds.
US, Canada · ▲ Bullish · 10 Aug 01:00 UTC · Trade press
Zayo and NVIDIA announced a partnership to expand network capacity supporting AI data centres and GPU clusters across North American sites.
Why this matters: Adds dedicated high-bandwidth interconnect for distributed GPU training and inference, reducing latency and congestion in AI cluster networks.
South Korea · ▲ Bullish · 10 Aug 06:00 UTC · Trade press
Samsung's HBM4 yield has reached approximately 80%, intensifying the production race with SK Hynix as both suppliers ramp capacity for AI accelerators.
Why this matters: Higher Samsung yields accelerate HBM4 supply availability and could compress memory pricing, reducing a key cost component in GPU cluster builds.
US · ▲ Bullish · 10 Aug 16:25 UTC · Company/PR
NVIDIA released Magpie Multilingual TTS, a 364M-parameter open-weights model supporting 12 languages, achieving 32ms time-to-first-audio on B200 GPUs and 320× real-time throughput at 64 concurrent streams.
Why this matters: B200 inference performance for latency-sensitive voice workloads, setting a benchmark for real-time agentic voice applications and competing inference platforms.
China · ▲ Bullish · 10 Aug 07:22 UTC · Trade press
Moore Threads reported H1 2026 revenue growth of 147% year-over-year and is planning a Hong Kong stock exchange listing.
Why this matters: Accelerating domestic GPU adoption in China and Moore Threads' competitive positioning against NVIDIA in the restricted market.
US · ◆ Neutral · 10 Aug 00:00 UTC · Company/PR
DESI released a 5.6-trillion-pixel 2D map of the universe containing 4 billion celestial objects, requiring approximately 100,000 GPU node-hours and 15,000 CPU node-hours on NERSC's Perlmutter supercomputer.
Why this matters: GPU-accelerated scientific computing at scale and GPU-enabled processing pipelines for large-scale data reduction workflows.
US · ◆ Neutral · 10 Aug 18:17 UTC · Company/PR
Dell announced Muse Glimmer, a 30B-parameter AI model, can run locally on Dell Pro Precision hardware without cloud connectivity.
Global · ▲ Bullish · 10 Aug 03:45 UTC
AP Memory, a memory supplier, appointed Jian Zhang, former Qualcomm vice president, to its global leadership team, company-stated.
Why this matters: AP Memory's ambitions in high-performance memory markets; Qualcomm background suggests focus on custom silicon and HBM integration for AI accelerators.
US · ◆ Neutral · 10 Aug 16:00 UTC · Company/PR
HYPERSCALER · 5
China · ▼ Bearish · 10 Aug 02:08 UTC · Trade press
China deployed its first 100,000-card domestic AI super-cluster using domestic accelerators, marking a milestone in the national compute grid buildout.
Why this matters: China's progress toward compute self-sufficiency and reduces reliance on NVIDIA exports, reshaping the global GPU market dynamics.
US · ▲ Bullish · 10 Aug 15:45 UTC · Company/PR
Amazon Bedrock now integrates OpenAI models (GPT-5.4, GPT-5.5, GPT-5.6) with web search capability, and introduced AgentCore Runtime Instances for deploying AI agents on dedicated infrastructure with predictable performance.
Why this matters: Expands AWS's inference endpoint portfolio to include OpenAI models and dedicated agent runtime, competing with neocloud inference providers and hyperscaler investment in agentic infrastructure.
US · ▲ Bullish · 10 Aug 16:00 UTC · Company/PR
Malachyte deployed transformer-based neural networks on Google Compute Engine completing recommendation inference in under 100 milliseconds, helping retailers double and triple sales through real-time personalisation.
Why this matters: Hyperscaler GPU inference infrastructure supporting sub-100ms latency for production agentic retail systems, GCP's inference-per-dollar positioning.
US · ◆ Neutral · 10 Aug 16:00 UTC · Company/PR
Google introduced ClusterNetworkPolicy (CNP) to GKE, an open-source Kubernetes standard with hierarchical tier system for cluster-wide network security, currently in preview in GKE 1.36+.
US · ◆ Neutral · 10 Aug 15:51 UTC · Company/PR
Amazon OpenSearch Serverless increased collection limit from 1,500 to 10,000 per collection group, reducing per-collection cost for multi-tenant applications by avoiding separate OCU provisioning.
DATA CENTRE · 1
US · ◆ Neutral · 10 Aug 11:00 UTC · Company/PR
Compass Datacenters, a Dallas-based data centre developer, publicly supported Texas Governor Abbott's call to raise development standards on 10 August 2026, company-stated.
ENERGY & POWER · 16
US · ▼ Bearish · 10 Aug 15:45 UTC · Trade press
PJM Interconnection set an all-time peak demand record above 168,000 MW on July 2, 2026, requiring federal emergency orders including forced data centre curtailment, with day-ahead prices exceeding $2,000/MWh in parts of the footprint.
Why this matters: Acute grid stress from data centre load growth, forcing emergency curtailment and tripling electricity prices year-over-year, raising interconnection risk and capex requirements for new data centre sites.
US · ▲ Bullish · 10 Aug 13:27 UTC · Trade press
Base Power closed a Series D funding round valuing the company at $13 billion, with plans to expand manufacturing from 40 MW monthly to 100 MW by year-end, and is pursuing 'bring your own capacity' partnerships with data centre developers.
Why this matters: Emergence of distributed battery-as-capacity model for data centres, with Base operating as vertically integrated energy provider in Texas; Google's 100 MW BYOC deal with Voltus exemplifies the trend.
Europe · ▼ Bearish · 10 Aug 13:28 UTC · Trade press
Seven European hydro systems produced roughly 13 TWh less hydro between April and July 2026 than the 2019-25 average, a shortfall of one fifth, with Rhine discharge at 37% of norm and Hungarian nuclear output falling 45% in three days.
Why this matters: Power supply constraint in Europe from drought and nuclear cooling limits, raising electricity prices (€117-122/MWh in July vs €88-103 a year prior) and increasing capex risk for data centre builds in the region.
China · ▲ Bullish · 10 Aug 11:59 UTC · Trade press
Changjiang unit 4 control room commissioned 31 July; units 3 and 4 (two Hualong One reactors) scheduled fully operational in early 2027 with CNY40 billion total investment.
Why this matters: Two new 1 GW+ reactors coming online in Hainan Province add baseload capacity for data centre loads in China's southern AI hub, reducing reliance on coal and grid constraints.
Global · ▼ Bearish · 10 Aug 07:30 UTC · Newsroom
Academic analysis suggests data centre capacity constraints and power limitations could dampen AI infrastructure earnings growth.
Why this matters: Flags power and cooling as binding constraints on AI capex returns, concerns about grid interconnection queues and PPA availability.
US · ▲ Bullish · 10 Aug 11:00 UTC · Newsroom
Fortune reports on SpaceX lockup expiry, Amazon's power plant acquisition, and Google's AI maps feature rollback.
Why this matters: Amazon's power plant acquisition hyperscaler commitment to securing on-site generation for data centre loads; relevant to the energy constraint on GPU cluster expansion.
Slovakia · ▲ Bullish · 10 Aug 13:14 UTC · Trade press
Mochovce 4 achieved first criticality on August 6, 2026, marking the start of controlled chain reaction in Slovakia's fourth VVER-440 reactor with 471 MW capacity, providing 13% of Slovakia's electricity when operating at full capacity.
Why this matters: Adds 471 MW of nuclear baseload to European grid, supporting long-term data centre power supply in Central Europe, though commercial operation still pending 144-hour full-power demonstration run.
Iran · ▲ Bullish · 10 Aug 11:09 UTC · Trade press
Rosatom increased staff at Iran's Bushehr nuclear plant to 25 with plans to reach 100 by autumn; Bushehr unit 2 reactor assembly over 60% complete, targeting 2029 operation.
Why this matters: Two new VVER-1000 units at Bushehr contribute to Iran's 20 GW nuclear target by 2041, adding baseload capacity independent of grid constraints or fossil fuel imports.
US · ▲ Bullish · 10 Aug 14:31 UTC · Trade press
The MORE American Fuel Act would allow uranium enrichment facility construction before NRC licensing, driven by the US ban on Russian low-enriched uranium taking full effect in 2028.
Why this matters: Accelerates domestic uranium enrichment capacity, supporting long-term nuclear generation for data centre power supply, though timeline extends beyond 2028 for material capacity additions.
Finland · ▲ Bullish · 10 Aug 12:05 UTC · Trade press
Posiva's Onkalo repository received a favourable safety assessment from Finland's Radiation and Nuclear Safety Authority, clearing a major hurdle toward an operating licence by end-2026.
Why this matters: Operational spent fuel storage removes a regulatory bottleneck for nuclear power expansion, enabling long-term grid stability for data centre loads across Northern Europe.
Global · ◆ Neutral · 10 Aug 10:06 UTC · Trade press
Advanced sodium-ion battery prototypes now approach 200 Wh/kg energy density, up from current commercial range of 100-160 Wh/kg, with first-generation commercial batteries entering energy storage market.
Why this matters: Sodium-ion batteries offer lower cost and supply-chain resilience versus lithium for stationary data centre backup power, though cycle life and self-discharge remain barriers to large-scale adoption.
US · ◆ Neutral · 10 Aug 15:40 UTC · Trade press
Sunrun cut full-year 2026 cash generation forecasts after Q2 earnings and launched a distributed compute pilot placing high-powered chips in customer homes, with commercialisation unlikely before late 2027 or 2028.
Why this matters: Early-stage exploration of distributed GPU compute at the edge, residential solar and battery infrastructure, but timeline suggests no material revenue impact on GPU infrastructure in 2026-2027.
US · ◆ Neutral · 10 Aug 13:18 UTC · Trade press
Avantus closed a US$1.05 billion corporate credit facility to fund solar and battery storage development across California and the Desert Southwest, with a 24 GW development pipeline comprising 13 GW solar paired with 44 GWh battery storage.
Why this matters: Continued capital availability for renewable-plus-storage projects, though no explicit data centre offtake agreements disclosed; relevant to long-term power supply for western US data centre expansion.
Europe · ◆ Neutral · 10 Aug 07:57 UTC · Trade press
European transmission system operators forecast a 2 GW reduction in photovoltaic generation during a solar eclipse on 12 August against 125 GW installed capacity, with France alone losing around 1,800 MW.
Why this matters: Grid vulnerability to renewable intermittency, reinforcing the need for dispatchable generation or storage to support data centre loads in renewable-heavy grids.
Italy · ◆ Neutral · 10 Aug 09:49 UTC · Trade press
Italy's Ministry of Environment and Energy Security set a 10 GW capacity quota for photovoltaic plants under the new FerX incentive scheme with reference strike price of €80/MWh.
Why this matters: Large-scale PV deployment in Southern Europe adds renewable capacity to grids feeding data centre clusters, though intermittency requires paired storage or dispatchable generation.
Brazil · ◆ Neutral · 10 Aug 08:02 UTC · Trade press
Neoenergia will invest $610 million from 2026 to 2030 to modernise Brasília's distribution network, adding five new substations.
Why this matters: Grid modernisation in Brazil's capital supports future data centre expansion, though no AI compute load is explicitly mentioned.
REGULATION & POLICY · 3
US · ▲ Bullish · 10 Aug 13:00 UTC · Company/PR
Cloudflare achieved FedRAMP Class D (High) certification, up from FedRAMP Moderate in 2022, enabling it to handle the nation's most sensitive unclassified data, and is pursuing US Department of Defence Impact Level 4 authorisation.
Why this matters: Positions Cloudflare's global network for US government and defence AI workloads, competing with hyperscaler-hosted classified compute and demand for independent infrastructure providers in the defence sector.
US · ▲ Bullish · 4 Aug 17:15 UTC · 2 sources
The National Science Foundation launched new State and Regional AI Infrastructure Hubs to support AI-enabled scientific research across the US.
Why this matters: Government-funded GPU infrastructure programme may create new demand for compute capacity and establish regional AI compute standards.
US · ◆ Neutral · 10 Aug 14:00 UTC · Company/PR
OpenAI sent a letter to Texas Governor Greg Abbott outlining commitment to responsible AI infrastructure growth that benefits Texans.