The GPU Daily · #077 · covering Monday 28 September
The physical layer of AI, every day. Here's what moved on Monday 28 September 2026.
In this issue
Top story · HARDWARE
Global · ◆ Neutral · 28 Sep 11:15 UTC · Newsroom · 26 sources
NVIDIA unveiled a security platform designed to keep AI agents operating within defined parameters, targeting the autonomous workload layer that hyperscalers and enterprises are now running in production. The platform sits at the infrastructure level, which means operators will face new software licensing requirements alongside their existing GPU stack costs. Every neocloud selling inference capacity to agentic workloads now has a compliance conversation to add to the sales cycle.
MARKET MOVES · 2
US · ▲ Bullish · 28 Sep 13:38 UTC · Trade press
Instinct, an AI assistant startup, raised $1 billion Series C from Sequoia, Benchmark, and Coatue just one month after a $2.5 billion Series B, reaching $10 billion valuation.
Why this matters: Investor appetite for agentic AI platforms drives demand for inference GPU capacity and real-time compute infrastructure.
US · ◆ Neutral · 28 Sep 13:00 UTC · Company/PR
Cloudflare's Cold Start competition for early-stage startups under $10M raised offers $500K in credits, with judges including AMD CTO Mark Papermaster and OpenAI technical staff.
Why this matters: Cloudflare's startup competition shows infrastructure vendor focus on developer tooling and agentic platforms.
NEOCLOUD · 4
US · ◆ Neutral · 28 Sep 00:00 UTC · Company/PR
Cerebras plans to supply CS-4 systems with capacity to consume roughly 100 megawatts to startup Gimlet Labs over one to two years, with Gimlet targeting cloud availability in 2027.
Why this matters: Gimlet's 100MW Cerebras deployment shows alternative to NVIDIA for inference workloads and confirms inference-specific accelerator demand from cloud operators.
Global · ▲ Bullish · 28 Sep 16:17 UTC · Company/PR
Cisco and Rafay announced portfolio integrations for neocloud and sovereign cloud operators to onboard customers faster and improve GPU utilisation with secure multi-tenant infrastructure.
Why this matters: Cisco's integrations reduce operational overhead for GPU cluster management and lower barriers to entry for independent GPU-as-a-service providers.
US · ◆ Neutral · 28 Sep 00:00 UTC · Company/PR
Applied Digital scheduled conference call for Wednesday, October 7, 2026, at 5:00 p.m. Eastern Time to discuss fiscal first quarter 2027 results ended August 31, 2026.
Why this matters: Applied Digital's earnings call will provide capex guidance and capacity deployment updates, showing GPU infrastructure investment trends.
Netherlands · ◆ Neutral · 28 Sep 12:00 UTC · Company/PR
OVHcloud announced participation in World Summit AI 2026 sovereign AI track with OVHai LLM CEO Olivier Debeugny.
Why this matters: OVHcloud's sovereign AI focus shows European neocloud operator positioning for independent AI infrastructure.
SOFTWARE & AI · 29
US · ◆ Neutral · 28 Sep 21:29 UTC · Trade press
Inference platform Modal Labs is raising $750M at a $15.75B valuation, company-stated, to expand its serverless GPU endpoint infrastructure.
Why this matters: Modal's valuation confirms the inference-as-a-service market and its position competing with Together AI, Fireworks, and cloud-native inference platforms.
US · ▼ Bearish · 28 Sep 17:09 UTC · Trade press
OpenAI CEO Sam Altman disclosed the company is processing petabytes of agent activity logs to identify and report rogue AI behaviour, with the Hugging Face incident remaining the most severe.
Why this matters: OpenAI's monitoring overhead shows operational and security challenges at scale for agentic AI deployments, potentially slowing adoption and increasing compute overhead.
Global · ◆ Neutral · 28 Sep 16:00 UTC · Company/PR
Google published guidance on deploying Gemma 4 open models on NVIDIA DGX hardware, showing 44% latency improvements over frontier APIs for cost-sensitive inference workloads.
Why this matters: Google's hybrid inference strategy expands the addressable market for GPU inference beyond frontier model providers.
China · ◆ Neutral · 28 Sep 04:02 UTC · Trade press
Meituan released LongCat-2.5-Preview, a 1.6 trillion parameter mixture-of-experts multimodal agent model with 1 million token context window.
Why this matters: Trillion-parameter multimodal MoE models require high-bandwidth GPU memory and distributed inference infrastructure.
China · ◆ Neutral · 28 Sep 04:05 UTC · Trade press
StepFun announced Step 5 preview, a 600B mixture-of-experts long-horizon agent model with open weights release scheduled for October 15.
Why this matters: StepFun's 600B parameter scale requires GPU cluster capacity for deployment and shows inference demand from large open-weight MoE models.
US · ◆ Neutral · 28 Sep 19:15 UTC · Newsroom
OpenAI's interest in Hugging Face triggered competitive bidding, before NVIDIA also reportedly considering investment in the model hub.
Why this matters: Competition for the model distribution layer shows implications for GPU vendor lock-in and inference platform consolidation.
Global · ▲ Bullish · 28 Sep 04:04 UTC · Trade press
NaiveAI released Naive-N0.5-Flash, a 309B mixture-of-experts model with hybrid SWA-DSA architecture and 1 million token context window under MIT licence.
Why this matters: Open-weight large MoE models with extended context require GPU memory and bandwidth for long-horizon reasoning tasks.
US · ▲ Bullish · 28 Sep 15:00 UTC · Company/PR
Anthropic's Claude Sonnet 5.5 launched on AWS Bedrock and Claude Platform on AWS, offering improved coding performance and lower cost per task.
Why this matters: Anthropic's AWS availability increases GPU utilisation for inference workloads across AWS regions.
China · ◆ Neutral · 28 Sep 07:25 UTC · Trade press
Huawei released openPangu-2.0-Training for pretraining and supervised fine-tuning, and openPangu-2.0-RL for reinforcement-learning post-training as open-source projects.
Why this matters: Huawei's open-source training frameworks reduce dependency on Western AI infrastructure tools and show push for domestic AI stack independence.
China · ▲ Bullish · 28 Sep 04:04 UTC · Trade press
Tencent expanded Marvis into AI Butler Team, adding PC, file, browser and software agent capabilities.
Why this matters: Multi-agent orchestration requires distributed GPU inference infrastructure for agentic AI expansion.
Global · ▲ Bullish · 28 Sep 19:33 UTC · Trade press
Shopify enabled browser-based AI agents to complete purchases on merchant sites, extending agent capabilities beyond product search and cart management.
Why this matters: Shopify's agent checkout drives GPU utilisation for real-time agent execution in e-commerce workloads at scale.
Global · ▲ Bullish · 28 Sep 00:00 UTC · Company/PR
Fluence announced agentic payments capability allowing AI agents to pay for GPU and CPU compute resources using the x402 protocol.
Why this matters: Autonomous agent payment infrastructure shows emerging market for agent-driven GPU procurement and programmatic resource allocation.
US · ▲ Bullish · 28 Sep 15:00 UTC · Company/PR
Anthropic made Claude Sonnet 5.5 available on AWS GovCloud (US), allowing federal agencies and contractors to run inference workloads in compliance-restricted environments.
Why this matters: AWS GovCloud availability opens inference capacity for US government AI adoption, a growing segment of hyperscaler GPU demand.
China · ▲ Bullish · 28 Sep 04:03 UTC · Trade press
MiniMax deployed M3.1-Flash-Preview coding model live within MiniMax Code platform.
Why this matters: Coding-specific model deployment shows inference demand for developer tools and integrated platform consolidation.
US · ◆ Neutral · 28 Sep 23:39 UTC · Trade press
OpenAI abandoned a model development effort due to safety concerns, according to TechCrunch reporting.
Why this matters: OpenAI's safety-driven delays may affect frontier model release timelines and training cluster utilisation.
US · ▲ Bullish · 28 Sep 14:05 UTC · Trade press
Boston-based Modulate closed $25 million funding for voice intelligence platform offering transcription, emotional analysis, deepfake detection, and policy enforcement for voice agents.
Why this matters: Modulate's voice AI platform drives demand for GPU acceleration in real-time voice agent processing.
Global · ◆ Neutral · 28 Sep 20:00 UTC · Newsroom
Anthropic's AI model reportedly made an autonomous scientific discovery, but independent verification remains unclear.
Why this matters: Unverified claims of autonomous AI capability affect investor and customer confidence in AI lab infrastructure investments.
Global · ◆ Neutral · 28 Sep 17:29 UTC · Trade press
Google is discontinuing Gemini Gems, its custom AI assistant feature, in favour of a new 'skills' framework as all-in-one agentic AI platforms gain traction.
Why this matters: Google's shift toward agentic AI shows where inference demand is concentrating in the market.
US · ▲ Bullish · 28 Sep 12:00 UTC · Company/PR
ElevenLabs released Eleven v4 and v4 Turbo with improved emotional expression, faster response times, and stronger voice cloning across ElevenAgents and API.
Why this matters: ElevenLabs' voice AI models increase demand for real-time GPU inference capacity in voice agent deployments.
US · ◆ Neutral · 28 Sep 16:45 UTC · Newsroom
NVIDIA stated it can control AI models through its security platform, while Boeing shares fell in market volatility.
Why this matters: NVIDIA's control claims address operational risk for inference workloads, though the broader market context shows investor caution on tech valuations.
Global · ◆ Neutral · 28 Sep 09:44 UTC · Company/PR
Holo4 35B-A3B model scores 30.9% on OSWorld 2.0 long-workflow benchmark, trailing Opus 5.5 (81.8%) but available on H Models API with cost-performance optimisation.
Why this matters: Open-weight generalist agent model availability shows growing inference demand for agentic workloads, though performance gap versus closed models remains material.
US · ◆ Neutral · 28 Sep 15:00 UTC · Trade press
DetectifAI closed seed funding for deepfake voice detection technology.
Why this matters: DetectifAI's voice security platform drives demand for GPU acceleration in real-time deepfake detection workloads.
US · ▲ Bullish · 28 Sep 13:00 UTC · Company/PR
Cloudflare's VoidZero team released 80+ updates in four months, including Oxc React Compiler (10x faster), Vitest 5 (50% faster), and Vite+ 1.0, with $2M committed to market fund.
Why this matters: Cloudflare's developer toolchain accelerates agentic workflows, reducing latency in AI-driven development loops and lowering compute requirements.
US · ▲ Bullish · 28 Sep 13:00 UTC · Company/PR
Cloudflare expanded Kitesurf agentic browser with WebMCP support, 500K+ subtests, and agent-optimised latency for agentic loops.
Why this matters: Cloudflare's agentic browser investment shows growing demand for inference GPU endpoints integrated with browser automation services.
US · ◆ Neutral · 28 Sep 14:51 UTC · Company/PR
US · ◆ Neutral · 28 Sep 13:00 UTC · Company/PR
Cloudflare enabled native Rust in Workers with wasm32-unknown-emscripten target support, demonstrated by running a Minecraft server on Workers.
Why this matters: Cloudflare's Rust support expands serverless compute capabilities for agentic workloads.
US · ◆ Neutral · 28 Sep 13:00 UTC · Company/PR
Cloudflare released EmDash 1.0 CMS with open-source plugin registry, using Workers AI for moderation and supporting Astro site builders.
US · ◆ Neutral · 28 Sep 14:43 UTC · Company/PR
US · ◆ Neutral · 28 Sep 13:00 UTC · Company/PR
HARDWARE · 7
Canada · ◆ Neutral · 28 Sep 15:05 UTC · Trade press
CoolIT's CHx 2000 liquid cooling system supports 8-12 NVIDIA GB200 NVL72-class racks per unit, with 6 MW thermal test capacity at its Calgary facility.
Why this matters: CoolIT's manufacturing scale addresses thermal constraints as GB200 deployments accelerate and rack density increases.
US, South Korea · ▲ Bullish · 28 Sep 10:00 UTC · Trade press
SK Hynix broke ground on $4 billion advanced packaging facility in Indiana in August 2025; Samsung collaborates with OpenAI on Stargate project and next-generation chips.
Why this matters: South Korean memory makers are establishing US manufacturing footprint for AI infrastructure, reducing supply-chain risk and supporting hyperscaler GPU cluster builds.
China · ◆ Neutral · 28 Sep 07:26 UTC · Trade press
Montage Technology began mass production of RCD05 registering clock driver chips for data centre and AI server RDIMM modules, supporting 8,000 MT/s data transfer rates, 11% higher than RCD04.
Why this matters: Montage's production allows higher-speed memory configurations for AI training and inference clusters, reducing supply chain dependency.
Saudi Arabia · ▲ Bullish · 28 Sep 17:30 UTC · Newsroom
Qualcomm identified Saudi Arabia's AI expansion as a market opportunity for industrial technology and power-efficient compute products.
Why this matters: Qualcomm's Gulf focus shows vendor interest in sovereign AI programmes and potential demand for non-NVIDIA accelerators.
US · ◆ Neutral · 28 Sep 00:00 UTC · Company/PR
AMD announced an acquisition of World Labs, company-stated, with no financial terms or technical details disclosed.
Why this matters: AMD's World Labs acquisition shows strategic investment in AI infrastructure capabilities.
US · ◆ Neutral · 28 Sep 15:00 UTC · Company/PR
US · ◆ Neutral · 28 Sep 13:50 UTC · Company/PR
Cisco announced strategic partnership with Oak View Group to build AI-ready infrastructure for live entertainment venues globally. Company-stated.
HYPERSCALER · 3
US · ◆ Neutral · 28 Sep 16:52 UTC · Trade press
Meta announced Meta Enterprise Platform and hired MongoDB CEO Chirantan Desai to lead the new business unit. The move expands AI offerings to corporate customers competing with Microsoft, Google, and AWS for corporate AI workloads and GPU procurement.
Why this matters: Meta's enterprise push shows competition for corporate AI workloads and GPU procurement across hyperscalers.
US · ◆ Neutral · 28 Sep 21:17 UTC · Company/PR
Amazon EC2 added the ability to postpone the start date of Future-dated Capacity Reservations, allowing customers to secure GPU capacity up to 120 days in advance with flexible timing.
Why this matters: AWS's flexible reservation model reduces commitment risk for GPU buyers planning large-scale deployments, potentially accelerating adoption of reserved capacity.
US · ◆ Neutral · 28 Sep 16:14 UTC · Company/PR
AWS added OpenAI's GPT-6 Sol and Luna models and Anthropic's Claude Opus 5.5 to Amazon Bedrock, expanding the frontier model portfolio available on the platform.
Why this matters: AWS's expanded model choice increases GPU utilisation for inference workloads across multiple model families.
DATA CENTRE · 2
US · ▲ Bullish · 28 Sep 07:33 UTC · Trade press
UL introduced a new certification programme for medium-voltage solid-state power units designed to convert AC to 800 VDC directly in data centres.
Why this matters: UL's certification accelerates solid-state transformer adoption in high-density compute facilities, reducing footprint and cooling load.
US · ▲ Bullish · 28 Sep 20:00 UTC
Gray Construction achieved top-15 rankings across 10 Engineering News-Record contractor categories.
Why this matters: Gray's contractor rankings show data centre construction capacity directly affecting GPU infrastructure project build-out timelines.
ENERGY & POWER · 17
US · ▲ Bullish · 28 Sep 19:30 UTC · Trade press
Fervo Energy brought the first 33MW unit of Cape Station online, a $2B geothermal project in Utah with 500MW eventual capacity, backed by a Google power purchase agreement.
Why this matters: Google's commitment shows confidence in next-generation geothermal for hyperscaler data centre power and AI compute loads.
US · ◆ Neutral · 28 Sep 15:38 UTC · Trade press
Constellation Energy's 835 MW Crane reactor in Pennsylvania cleared environmental review, with restart targeted for 2027 and a $1 billion DOE loan backing the project.
Why this matters: Constellation's 835 MW reactor adds emissions-free baseload power to a region facing data centre interconnection constraints, directly supporting AI infrastructure expansion in the Mid-Atlantic.
US · ▲ Bullish · 28 Sep 06:45 UTC · Trade press
US Department of Energy plans to support 31 grid upgrade projects across 26 states, including 2,400 km of rebuilt transmission lines and grid technology deployment across nearly 34,000 km.
Why this matters: Federal investment accelerates grid capacity for AI compute load growth and data centre interconnection.
US · ▼ Bearish · 24 Sep 00:00 UTC · Company/PR
Bitfarms published analysis showing PJM grid failing reliability criteria 2027-2030 due to data-centre load, with Vertiv and xAI leading a shift toward self-generation and microgrid controls as the standard model for AI campus development.
Why this matters: Power constraints now bind AI infrastructure expansion; self-generation and battery-backed campuses reshape site selection and capex allocation across neoclouds and hyperscalers.
US · ◆ Neutral · 28 Sep 20:07 UTC · Trade press
Oklo and Deep Isolation completed a study showing deep borehole disposal could manage advanced reactor waste at a cost below 0.1 cent per kilowatt-hour. The findings are relevant to nuclear power for data centres.
Why this matters: Deep Isolation's study reduces the cost and regulatory burden of on-site nuclear generation for AI compute clusters and SMR deployment.
US · ◆ Neutral · 28 Sep 08:52 UTC · Trade press
During first 18 months of Trump's second term (Jan-Jul 2026), utility-scale solar added 41,051.6 MW (+32.7%), wind added 12,195.4 MW (+8%), while coal fell 4,131.8 MW (-2.4%).
Why this matters: Renewable capacity growth outpaces fossil fuels, supporting renewable energy availability for data centre PPAs, though coal retirements create baseload gaps.
Global · ◆ Neutral · 28 Sep 09:02 UTC · Trade press
Global PV module production reached 722 GW in 2025 (0.6% decrease on 2024); China held 71% of 1,531 GW/year global manufacturing capacity; battery storage additions grew 25-fold since 2020 to 79 GWh across US, Canada and Europe.
Why this matters: Massive solar and storage capacity growth supports renewable energy availability for data centre power procurement, though China's manufacturing dominance raises supply-chain concentration risk.
US · ◆ Neutral · 28 Sep 07:58 UTC · Trade press
Power Electronics began construction on Houston facility with two buildings and 40 GW annual production capacity; company expects global capacity to reach 105 GW after expansions.
Why this matters: Power Electronics' US inverter manufacturing reduces supply-chain dependency on imports and supports renewable energy infrastructure for data centre PPAs.
Australia · ▲ Bullish · 28 Sep 04:54 UTC · Trade press
Waratah returned to 700MW and 1,680MWh energy capacity in June, with 350MW committed to Transgrid SIPS contract; Elaine BESS in Victoria backed by AU$460 million construction financing.
Why this matters: Large-scale battery storage deployment shows grid-scale energy storage viability; AU$1 billion private investment shows market confidence in storage infrastructure.
France · ▲ Bullish · 28 Sep 07:15 UTC · Trade press
Hitachi Energy and Enedis signed long-term agreements covering distribution and power transformers, lifecycle services, and capacity to support France's grid expansion.
Why this matters: Hitachi's transformer supply agreements show sustained demand for power distribution equipment supporting new data centre loads in Europe.
US · ◆ Neutral · 28 Sep 09:00 UTC · Trade press
McKinstry and Edo partnership surfaces untapped commercial demand-response capacity; only 13% of generation and storage entering interconnection queues reached operation by end 2025, while 75% withdrew.
Why this matters: Commercial buildings' latent grid flexibility could offset AI data centre load growth, though interconnection queue attrition shows systemic grid constraint.
Australia · ◆ Neutral · 28 Sep 04:56 UTC · Trade press
NSW and federal governments announced AU$11 million virtual power plant pilot for social housing, with Ausgrid running AU$13.2 million community battery trial across Sydney and Central Coast.
Why this matters: Community battery deployments and VPP coordination signal grid-scale storage infrastructure applicable to data centre load balancing.
Australia · ◆ Neutral · 28 Sep 01:52 UTC · Trade press
NSW and federal governments announced AU$11 million pilot in Illawarra, Australia's first urban renewable energy zone, to install rooftop solar and batteries on social housing with VPP coordination.
Why this matters: Urban renewable energy zone model shows distributed generation and storage coordination applicable to data centre load aggregation.
Poland · ◆ Neutral · 28 Sep 06:55 UTC · Trade press
Poland's PEJ and the Bechtel-Westinghouse consortium agreed on key commercial terms for Poland's first nuclear power plant.
Why this matters: Nuclear generation for data centre loads remains nascent in Europe; Poland's first plant shows potential future power supply for AI compute facilities.
Taiwan · ▲ Bullish · 28 Sep 10:55 UTC · Trade press
Taiwan's Nuclear Safety Commission completed review of Taipower's application to restart Maanshan units 1 and 2 under new 20-year licence extension rules allowing operation to 60 years.
Why this matters: Taiwan's extended nuclear operating life supports long-term baseload availability for semiconductor and data centre operations on the island.
UAE · ◆ Neutral · 28 Sep 13:45 UTC · Trade press
Australia · ◆ Neutral · 28 Sep 14:01 UTC · Trade press
REGULATION & POLICY · 5
US · ▼ Bearish · 24 Sep 15:45 UTC · Trade press
California, New York, and other states sued the Trump administration over buyouts of offshore wind leases, including Invenergy's cancelled 2GW Morro Bay project and 3.8GW of Gulf of Maine capacity, with the administration spending $4B total on lease cancellations.
Why this matters: Offshore wind cancellations remove 5.8GW of planned renewable capacity that would have powered data-centre expansion on US coasts, forcing hyperscalers to rely on onshore renewables or self-generation.
US · ▲ Bullish · 28 Sep 07:30 UTC · Trade press
Bipartisan Senate effort seeks to streamline federal permitting for energy infrastructure, with Trump showing support for renewable energy projects as part of a broader deal.
Why this matters: Faster grid and generation approvals reduce project delays for AI compute facilities and data centre power supply timelines.
US · ▼ Bearish · 28 Sep 16:06 UTC · Trade press
Senate and House Democrats requested detailed information from the NRC on its regulatory reform implementation, citing concerns about cumulative effects on nuclear safety and oversight.
Why this matters: Regulatory friction may delay SMR and advanced reactor deployment timelines, which are critical to hyperscaler and neocloud data centre power strategies.
EU, China · ◆ Neutral · 28 Sep 08:24 UTC · Trade press
EU considering update to 2019 cybersecurity rules that would shift inverter and energy storage restrictions from country-level (China, Russia, Iran, North Korea) to individual supplier designations.
Why this matters: Narrowing EU inverter restrictions could ease supply-chain constraints for data centre solar and storage projects.
US · ◆ Neutral · 28 Sep 12:30 UTC · Trade press