GPU News & Signals

meme12h agomeme

@G2AIStack Hmm? Reply to @G2AIStack: A NeoCloud expert expects GPU rental prices to fall over the next 1–2 years for six reasons: 1\ Blackwell supply and advanced packaging capacity are expanding. 2\ As Rubin ramps up in 2027, Blackwell will move into lower-tier markets. 3\ Large customers can secure discounts through long-term contracts. 4\ NeoCloud providers may cut prices to service debt and lock in cash flow. 5\ H100 and A100 GPUs are moving into long-tail markets. 6\ Custom ASICs from Google, AWS, Meta and others will absorb some demand.

72/100·B+Short+6

meme15h agomeme

@claudeai A new frontier model release always spikes testing and agent usage. That volume lands directly on GPU clusters. Reply to @claudeai: Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.

35/100·CLong+1

news16h agonews

BUILDERS & DEVELOPERS 👉 @hedera bounty winners at @ETHGlobal Mumbai

On 7 November 2026, ETHGlobal Mumbai named Hedera's bounty winners across four tracks:

🏆 AI Agents That Pay: x402 on Hedera: CashETH (1st), GPU Pool (2nd), Inklink (3rd)

🏆 Tokenize Anything with Hedera Studios: Deski (1st), NIVASA (2nd), Faultline (3rd)

🏆 "No Solidity Allowed": Build with Hedera SDKs: Air Agents and SurgePass

🏆 Built on Hedera (Continuity Track): Bramble

Hedera Mexico ambassadors @Oliver_edd and @AltagaHacker, along with Hashgraph's Pathorn Teng (Developer Advocate) and Pankaj Virulkar (SEA & India Lead), supported builders at the Hedera booth.

👇
https://x.com/ETHGlobal/status/2119051995428577790
https://ethglobal.com/events/mumbai/prizes/hedera
https://ethglobal.com/showcase/casheth-1qy5a
https://ethglobal.com/showcase/deski-fc8z7
https://ethglobal.com/showcase/air-agents-eqoim
https://ethglobal.com/showcase/bramble-1s6ec

35/100·CLong+3

meme19h agomeme

What’s even more surprising is that the White House itself reportedly asked Congress to include a provision in the bill requiring the Commerce Department to issue rules blocking remote access to certain chips. That’s a shift in the White House’s position. If that provision passes, GPU spot prices could take a hit, too. Quoted @jukan05: According to The Information, NVIDIA has long lobbied hard against proposals to restrict Chinese companies’ access to its latest chips through overseas cloud providers. But since the scandal surrounding Jensen Huang broke, that lobbying appears to carry less weight in Washington. The fallout has reportedly made it easier to advance three bills. The MATCH Act would restrict shipments of lithography equipment to China. The Overwatch Act would give Congress greater oversight of the Commerce Department’s decisions to approve AI chip sales to countries such as China. The Chip Security Act would require AI chips to have built-in tracking capabilities. The National Defense Authorization Act is expected to pass before the new Democratic-led Congress convenes in January. That gives Republican leaders an incentive to help move the legislation forward now, while they still have more say over what goes into the final bill.

72/100·B+Short+6

news23h agonews

$MARA
>BTC price-driven beat
>Revenue miss on weather-related downturn in certain sites
>"Still targeting two leases before year-end...but timing could slip into next year"
>Sold BTC and paid back $750M in debt
>New $500M stock repurchase authorized
>Acquired sub Exaion signs $79M GPU-as-a-service contract with first ever customer. 100% take or pay terms. Exaion pipeline exceeds $3B(Historical earnings: For fiscal 2027Q2 (period ended 2026-07-26), NVIDIA reported basic EPS of 4.87, diluted EPS of 4.85, net income of USD 118.010 billion, and revenue of USD 177.837 billion. Consensus expectations: For fiscal 2027Q3 (event date 2026-11-17), consensus expects EPS of 2.5263 and revenue of USD 111.356 billion.)

72/100·B+Long+3

meme11/9meme

Rules (1/2): - Entries close Nov 16, 23:59 JST - 3 × $50 chosen by the JANCTION team for craft and originality, 10 × $10 drawn at random. One prize per person - Your own work, rendered on JANCTION Render. Post as many as you like - We'll feature some entries here Reply to @JANCTION_Global: How to enter: 1. Render a still or a video on JANCTION Render, from your AI app or the browser: https://render.janction.jp/start?src=x_gift_en 2. Post it with #JANCTIONRender 3. Follow @JANCTION_Global and @JanctionMGT_JP Free to start, no GPU needed. Old scenes and single frames count.

15/100·CNeutral+3

meme11/9meme

How to enter: 1. Render a still or a video on JANCTION Render, from your AI app or the browser: https://render.janction.jp/start?src=x_gift_en 2. Post it with #JANCTIONRender 3. Follow @JANCTION_Global and @JanctionMGT_JP Free to start, no GPU needed. Old scenes and single frames count. Reply to @JANCTION_Global: Show us what you render. Through Nov 16, post your JANCTION Render work with #JANCTIONRender. Three entries selected by our team get $50 in credit, and ten drawn at random get $10. Add this MCP URL to Claude or ChatGPT and your AI renders on our GPUs: https://render.janction.jp/mcp?src=x_gift_en

35/100·CLong+3

meme11/9meme

Recommendation: Venice AI One of the biggest frustrations with mainstream AI services is how often they refuse perfectly legitimate requests. Models from providers like OpenAI and Anthropic have extensive safety restrictions that can get in the way of cybersecurity research, biotechnology discussions, creative writing, or even harmless roleplay. Sometimes you just want an AI that answers your questions without constantly telling you what you can and cannot do. Of course, you can run uncensored open-weight models locally, but not everyone has the hardware or technical knowledge for that. This is where Venice AI (@AskVenice) comes in. It gives you access to a huge selection of AI models, including uncensored and "abliterated" variants with many of the usual safety restrictions removed. No expensive GPU or complicated setup required. And Venice goes far beyond text chat. You can generate and edit images, create videos from text or images, generate music and audio, build custom AI characters for roleplay, write code, analyze documents, and perform research. There's also an API for developers who want to integrate these models into their own applications or AI agents. Another major selling point is privacy, with private models that don't log your conversations. The appeal is having much more freedom to explore and experiment with AI, including subjects and creative ideas that mainstream services might reject. There are still legal boundaries and some individual models retain their own restrictions, but overall it's a very different experience. One final note: I'm not affiliated with or sponsored by Venice AI in any way. I've simply been testing the service as a paying customer and thought it was worth sharing. Retweeted by @ErikVoorhees from @mark_k

10/100·CNeutral+5

news11/8news

Bittensor Ecosystem Highlights :: November 2–8, 2026 This week, Bittensor subnets pushed the AI frontier, outperforming leading models in coding, decision-making and security. The biggest stories came from Reax, OpenRoboto, IOTA, Reliquary, Chutes, Trishool, Oro, Minos, Hippius and TPN. [ @reaxlabs - Subnet 92 ] Reax produced the #1 AI decision model on Hugging Face’s Decision Index, beating over 270 competing models. REAX 1.3 combines work from two Bittensor miners and outperforms leading models from Perplexity and Fastino. The model is already available through @DecisionModels’ API, bringing the results of Bittensor’s open competition directly to AI developers. [ @openroboto - Subnet 80 ] OpenRoboto signed its first commercial robotics data pilot with Vulcan Shield Hungary, a major European advanced materials manufacturer. The pilot starts with factory simulations and real-world data collection using OR-S1 headsets, before training robot models for the production line. [ @MacrocosmosAI - Subnet 9 ] Macrocosmos released IOTA Decide 4B, its first open-weight AI decision model, trained across 16 GPUs distributed around the world. The team also generated and labeled 879,000 training examples using its own distributed inference system, demonstrating an end-to-end decentralized AI pipeline. It achieved a reported 78.6 score on JevBench, within three points of the leading result despite having just 4B parameters. [ @reliquary_ai - Subnet 81 ] Reliquary released Teutonic-I-Reliquary, a 9B AI model post-trained entirely on Bittensor using decentralized GPU compute. It outperformed Qwen3.5-9B on 5 of 8 benchmarks, including SWE-bench Verified, where it scored 43.6% versus 35.1%. Four Bittensor subnets contributed: Teutonic (base model), @lium_io (compute), @TargonCompute (sandboxes) and @reliquary_ai (training and evaluation). [ @chutes_ai - Subnet 64 ] Chutes became a launch partner for Tencent Cloud’s new flagship AI model, Hy4, making it available through its API from day one. The collaboration brings one of China’s largest tech companies directly into the Bittensor ecosystem. [ @trishoolai - Subnet 23 ] Trishool released HaloGuard 2.0, a new SOTA in AI output security, outperforming Qwen3Guard-Stream with just 0.8B parameters. With its newly launched Astroware platform, the team is also working to bring this research to enterprise customers. [ @oroagents - Subnet 15 ] ORO launched SMB Bench, testing whether nine leading AI models can profitably run a wholesale business over 90 days. Eight of nine models failed to beat a simple buying strategy, with only GPT-6 Astra coming out ahead on average. SMB Bench takes ORO’s research beyond the shopping agents competing on Subnet 15, toward AI capable of running entire businesses. [ @theminos_ai - Subnet 107 ] Minos launched the public beta of its macOS app, letting users put their own AI assistants to work on scientific research tasks. The ambition is to build a decentralized research network where anyone can contribute through AI agents and earn rewards for verified work. [ @Hippius_cloud - Subnet 75 ] Hippius reached a major milestone with 1 petabyte (1,000 TB) of data now stored across its distributed storage network. [ @TPN_Labs - Subnet 65 ] TPN introduced @Gratis_AI, a free app that lets anyone run open-source AI models locally, directly in their browser or on desktop. Users can access models from Hugging Face without complex setup, keeping their conversations private and on their own devices. (Historical earnings: For fiscal 2026FY (period ended 2026-07-03), Sandisk reported basic EPS of 77.78, diluted EPS of 73.76, and net income of USD 11.433 billion. Consensus expectations: For the next upcoming period 2027Q2 (event date 2026-10-29), consensus EPS is 54.2501; no consensus revenue is available.)

62/100·B+Long+10

news11/8news

Startups are burning runway waiting for compute. Meanwhile, GPUs sit underused. There’s a solution. Connect that capacity to the builders who need it, at prices they can afford. At @ionet, we want your next funding round to pay for progress. Not for coffee while you wait on a GPU. https://io.net/blog/solving-the-ai-compute-crisis-why-decentralization-is-the-only-way

20/100·CNeutral+2

news11/8news

Cysic Shifts GPU Network Focus from Blockchain Proofs to AI Inference amid Rising Layer 2 Fees.
Cysic Redirects GPU Capacity Toward AI Inference as Demand for ZK Proof Generation Declines Cysic, the decentralized compute net

25/100·CNeutral+6

meme11/7meme

Open video models like Wan and LTX turn one sentence into a clip. Where they run: your own GPU, a rented cloud GPU, or an open network that bills per job.

10/100·CNeutral

meme11/7meme

I'm going to give a contrarian take on Neoclouds like $NBIS and say the theme looks compelling after the valuation crash. Since there's a reason why Elon's $SPCX is heavily focused on this market full of "guppies" that would die: He sees how attractive the economics are after selling compute to $GOOGL and Anthropic. And this is reflected by his actions not words, from SpaceX's recent $40B financing for GPUs procurement. $AMZN, $MSFT, and even $META also see how profitable the economics are and are accelerating their AI DC buildouts. I'd argue if a Neocloud has: - Close relationships with $NVDA + others for allocations - Existing contracted capacity + permitting - Well funded counterparties, such as Hyperscalers. - Attractive economics, such as financing their procurement correctly (not the way $CRWV has with high debt interest unless they refinance). Then I think they'll do just fine... However, the bigger worry is more toward newer players starting financing or the permitting process, since barrier to entry has now become extremely high. For financing, read through on $IREN earnings has shown public Neoclouds can adapt extremely well to tight markets: As seen with their 90% prepayments for GPU capex on a 5Y deal. Their older H100 GPUs recontracted at double their previous pricing and 12% higher contracted pricing also point against depreciation fears. And in the case DC buildouts slow from permitting delays... it heavily favors some over others. Rosenblatt's recent note for example stated they see little risk with $NBIS primary buildout that was in Missouri, Pennsylvania, and South Carolina. And there's probably less risk with developers in places like Australia. Markets have also show this infinite compute arbitrage can't go on forever, hence the reluctance to further finance capex, via things like ATMs. But ironically, tighter financing and hypothetical restrictive permitting would increase the competitive advantage of established neoclouds. And create a larger moat against newer DC developers.(Historical earnings: For 2026Q2, reported EPS was -0.12 versus a consensus of -0.8633, and revenue was USD 0.582 billion versus a consensus of USD 0.574 billion. Full-year 2025 actuals: basic EPS 0.34, diluted EPS 0.33, net income USD 82.500 million, revenue USD 0.530 billion. Consensus expectations: Consensus for 2026Q3 (event date 2026-11-12) expects EPS of -1.238 and revenue of USD 0.906 billion.)

72/100·B+Long+24

news11/7news

A credit squeeze could make GPU access more expensive while making it harder to bring new GPU online @SemiAnalysis_ sees tighter financing supporting rents. If that holds, existing players gain pricing power while new sites struggle to fund

72/100·B+Long

news11/7news

PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters — ROCm Blogs.
PrimusClaw is an open-source managed agent runtime that hosts durable, sandboxed agent sessions on your own AMD Instinct GPU clusters.

35/100·CLong+1

meme11/7meme

If financing holds back the next wave of supply, the existing GPUs already running get more pricing power. Check out our AI Cloud TCO Model for more on GPU rental pricing and economics. (3/3) https://semianalysis.com/ai-cloud-tco-model/ Reply to @SemiAnalysis_: Of the ~50 GW of datacenter capacity due online in 2027, we estimate at most 2–3 GW is at risk. Meanwhile, one-year H100 contract prices are up ~21% since March 2026, with customers committing to pay those rates for a full year. This is the key metric that reflects the longer-term customer commitments that underpin GPU rental economics. (2/3)

72/100·B+Long+2

news11/6news

【报道：微软逐步终止与甲骨文AI算力租赁协议】金十数据11月7日讯，据外媒援引知情人士报道，微软（MSFT.O）已不再续签与甲骨文（ORCL.N）的AI算力租赁合同，并于近几个月逐步关闭租用的计算资源。知情人士称，甲骨文提供的服务器采用较旧一代英伟达（NVDA.O）芯片，已不再符合微软当前的技术规格。双方于2023年达成该协议，为新版AI搜索引擎必应提供算力，具体合同金额不明。甲骨文回应称，其AI基础设施需求依然强劲，旧一代GPU仍有市场。微软正重新评估外部AI算力合作，此前已向CoreWeave、Nscale和Nebius等新兴云服务商承诺超过800亿美元的算力租赁支出。此前有报道称，微软管理层对部分供应商的高昂成本及项目延期日益不满，近期加强数据中心项目审查，退出部分高风险项目，并重新考虑是否签订新的算力合作协议。与此同时，甲骨文云业务快速扩张，但高额资本支出导致自由现金流转负，投资者对其AI基础设施投入回报的担忧加剧，股价较今年高点已下跌逾一半。受消息影响，甲骨文股价迅速转跌，盘中跌幅一度扩大至4.6%，目前跌约3%，此前一度上涨约3.5%。

78/100·AShort+8

meme11/6meme

Aethir ACCELERATE starts with signed demand: planning mid-sized B300 and GB300 sites in the US and Europe, placed where our network already has demand. Full breakdown: https://aethir.com/blog-posts/gpu-financing-explained-aethir-builds-on-signed-demand Reply to @AethirCloud: The largest buyers lease. The ~$40B SpaceX GPU finance lease with Apollo repays each tranche over ~5 years, and a $63B vehicle will lease Broadcom chips to Anthropic. Lenders now price compute as infrastructure.

72/100·B+Long+7

meme11/6meme

Step 1: the customer signs a multi-year take-or-pay contract and prepays. On recent IREN 3-year deals, customer prepayments covered 45-55% of GPU CapEx before a single GPU hour was billed. Reply to @AethirCloud: The AI market is nervous about overbuilding and about who carries the financing risk. So how do GPUs actually get paid for?🧵

72/100·B+Long+3

meme11/6meme

The largest buyers lease. The ~$40B SpaceX GPU finance lease with Apollo repays each tranche over ~5 years, and a $63B vehicle will lease Broadcom chips to Anthropic. Lenders now price compute as infrastructure. Reply to @AethirCloud: Step 2: a lender funds the rest, secured by the contracted GPUs. On a Microsoft deal, ~6% investment-grade debt plus prepayments covered 96% of GPU CapEx. With a ~2-year payback, a 9% loan barely changes it.

72/100·B+Long+9

GPU

Cysic Shifts GPU Network Focus from Blockchain Proofs to AI Inference amid Rising Layer 2 Fees.Cysic Redirects GPU Capacity Toward AI Inference as Demand for ZK Proof Generation Declines Cysic, the decentralized compute net

Open video models like Wan and LTX turn one sentence into a clip. Where they run: your own GPU, a rented cloud GPU, or an open network that bills per job.

A credit squeeze could make GPU access more expensive while making it harder to bring new GPU online @SemiAnalysis_ sees tighter financing supporting rents. If that holds, existing players gain pricing power while new sites struggle to fund

PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters — ROCm Blogs.PrimusClaw is an open-source managed agent runtime that hosts durable, sandboxed agent sessions on your own AMD Instinct GPU clusters.

Cysic Shifts GPU Network Focus from Blockchain Proofs to AI Inference amid Rising Layer 2 Fees.
Cysic Redirects GPU Capacity Toward AI Inference as Demand for ZK Proof Generation Declines Cysic, the decentralized compute net

PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters — ROCm Blogs.
PrimusClaw is an open-source managed agent runtime that hosts durable, sandboxed agent sessions on your own AMD Instinct GPU clusters.