News Heat Trend
Project Overview
Inference is adistributed GPU cluster for LLM inference built on Solana. Inference.net is a global network of data centers serving fast, scalable, pay-per-token APIs for models like DeepSeek V3 and Llama 3.3.
Morgan Stanley: 70% of NVIDIA's FY28 Revenue Guidance Constrained by Supply; Inference Has Become the Largest and Continuously Expanding Data Center Segment
According to Chaohiang Research, JPMorgan's September 2 investor meeting minutes for NVIDIA indicate that NVIDIA is "comfortable" with its guidance for 70% year-over-year growth in FY28, and this guidance is supply-constrained rather than demand-constrained; with sufficient supply, the business could grow more than double. Inference has become the largest and expanding segment of the data center business. Eighteen months ago, inference and training each accounted for roughly half, but currently inference exceeds training and will continue to rise. Advanced wafers and memory are the two major supply bottlenecks, and NVIDIA maintains close cooperation with TSMC and the three major memory suppliers. The customer base continues to broaden, with OpenAI and Anthropic currently accounting for approximately 20% of end-user demand, which may approach 25% by FY28; new cloud providers now account for over 50% of AI computing infrastructure. Open-source and closed-source models will coexist, and gross margins for model developers are improving. NVIDIA supports long-term demand through revenue sharing, the PORTS-Pike campus, and a $500 billion private capital financing platform. Morgan Stanley maintains an Overweight rating with a price target of $320, based on approximately 20x the expected CY2026 EPS of $15.87.
Morgan Stanley: CoreWeave Adds Record 500MW Capacity in Single Quarter, High Debt and Customer Concentration Weigh on Valuation
According to TechFlow Research, Morgan Stanley's August Q2 earnings report indicated that CoreWeave added 500MW of net active power in a single quarter, exceeding any quarter in history, more than three times year-over-year. Management reaffirmed the target of reaching at least 8GW before 2030. FY26 revenue guidance midpoint was raised 2% to $12.4 billion to $13.2 billion, ARR midpoint was raised 3% to $18.5 billion to $19.5 billion. The company raised full-year capital expenditure guidance midpoint by 12% to $35.5 billion to $39.0 billion, Q3 capital expenditure guidance is $11.5 billion to $13.5 billion, higher than the market expectation of $10 billion. Managed Inference Platform (managed inference platform) ARR grew from $1 million to over $100 million, expected to reach at least $250 million by year-end. The research report judges that Q2 adjusted operating margin was about 8%, higher than expected, but Q3 margin guidance of 5.8% to 7.2% is lower than market expectations, Q4 margin needs to increase significantly to achieve full-year guidance. Morgan Stanley expects CoreWeave FY27 operating margin to be 15.9%, FY28 to be 22.4%, free cash flow to remain negative until 2028, and debt is expected to increase to approximately $38 billion by the end of 2026. Morgan Stanley maintains Equal-weight (in line with the market) rating and $99 price target.
AI Reasoning Startup General Compute Secures $400 Million Loan Backed by ASIC Chips for Inference
: AI reasoning cloud startup General Compute has obtained a $400 million loan from Upper90. This deal is the world’s first financing project to use dedicated inference chips as collateral. The company has built a proprietary AI reasoning cloud platform based on SambaNova’s self-developed ASIC chips, primarily targeting Agent-type AI computing workloads. Compared to traditional GPU clouds, it offers faster token processing speeds and lower operational latency. The hardware requires no water cooling and can be directly deployed in traditional data centers and idle cryptocurrency mining facilities.
DeepSeek Self-Develops AI Inference Chip, Plans to Break Dependence on NVIDIA and Huawei
According to Reuters, Chinese AI startup DeepSeek is developing its own AI chips, three informed sources revealed. The chip is designed specifically for inference scenarios, rather than for model training. The project was launched approximately one year ago and remains in the early stages. The company has engaged with chip design, wafer foundry, and storage enterprises, and has quietly increased recruitment of chip design engineers without publicly posting job listings. If successfully developed, DeepSeek will reduce its reliance on Nvidia and Huawei Ascend chips, following the trend of global AI giants such as OpenAI and Anthropic developing their own hardware. Affected by U.S. export controls, DeepSeek previously shifted from Nvidia H800 to Huawei chips. This self-developed chip is regarded as a significant strategic transformation. Meanwhile, DeepSeek also plans to complete its first round of external financing, with a fundraising scale of approximately $7 billion, and a valuation reaching $52 billion to $59 billion.
AMD middle managers accused of diverting multiple engineers to ATOM inference engine project with almost no customers
Odaily News: SemiAnalysis posted on X that AMD middle managers are spending the time of more than 20 engineers trying to expand their sphere of influence and secure promotions through a useless side project called the "ATOM Inference Engine," rather than allocating those resources to inference engines actually being used by real customers like Meta and xAI, such as SGLang and vLLM.SemiAnalysis said that dozens of AMD engineers, including principal engineers and AMD Fellows, have privately contacted the firm, stating that due to internal office politics at AMD and support from some middle managers, they are unable to discuss their concerns internally. Middle managers are consuming the time of more than 20 full-time engineers to develop this side project with almost no customers, and the engineers involved are all "10x engineers" who should have been dedicated to developing customer-facing inference engines like vLLM and SGLang.SemiAnalysis said that middle managers will likely respond by pointing to the only customer using the project in a production environment—Alibaba, based on Qwen—but further investigation shows that this is merely a limited deployment of ATOM by Alibaba's enterprise business unit, not the main business unit behind Qwen.SemiAnalysis stated that AMD management has diverted resources and limited internal GPU R&D clusters away from engineers supporting production-grade inference engines such as vLLM and SGLang, instead using them to expand their own influence and engage in office politics, showing a profound lack of respect for AMD engineers.
AMD launches AI programming platform Instinct Coder, which can reduce enterprise AI coding costs by 70%
Odaily News AMD, the semiconductor giant, announced the launch of its enterprise-grade AI programming platform, AMD Instinct Coder. The platform combines AMD chips, Supermicro servers, and Spectro Cloud software, aiming to help enterprises deploy AI coding assistants locally, reduce the cost of cloud-based AI models, and protect code and data security.AMD stated that Instinct Coder is an "out-of-the-box" end-to-end AI development platform, integrating AMD EPYC processors, AMD Instinct GPUs, Supermicro AI servers, Spectro Cloud PaletteAI Inference Launchpad software, and the AMD-optimized GLM-5.2 model. It can be used for software development scenarios such as code generation, application modernization, automated testing, and code review.AMD said that compared to relying on cutting-edge cloud-based AI models, Instinct Coder can help enterprises reduce total cost of ownership (TCO) by up to 70%, with the fastest payback period shortened to 6 months.AMD noted that more and more enterprises are looking to leverage AI to improve development efficiency, but face two major challenges: on one hand, the cost of invoking top-tier cloud models continues to rise; on the other hand, entrusting enterprise source code, intellectual property, and sensitive data to third-party services poses security and compliance risks.Through a local deployment model, Instinct Coder allows enterprises to maintain control over their data and code while providing more predictable infrastructure costs. The platform supports development tools such as Claude Code, OpenAI Codex, Visual Studio Code, and Cursor, with each node supporting up to 50 users (30 concurrent users).Additionally, the PaletteAI Inference Launchpad provided by Spectro Cloud enables AI workload management, model routing, request auditing, and cost monitoring, and supports invoking external models such as Anthropic, OpenAI, Google, or xAI when needed.AMD stated that Instinct Coder aims to help enterprises break free from the high costs of cloud-based AI services, accelerate AI-driven software development processes while ensuring data security and autonomous control.
Serenity: US AI Companies Lower Inference Costs to Counter the "Model Distillation" Challenge
"White-Haired Stock Guru" Serenity stated that while some observations in the UBS report hold anecdotal truth, the more noteworthy trend is the increasing number of Chinese-language reports regarding the distillation of Anthropic's models. Currently, many US startups and tech companies are opting to use cheaper Chinese models (such as DeepSeek) in their AI applications, as their unit task costs are significantly lower than those of inference models from Gemini, OpenAI, and Anthropic.Serenity believes this trend, driven by capitalism, creates a "typical paradox"—companies naturally gravitate towards lower-cost solutions, thereby eroding the leading advantage of US models. He proposes that the US needs to address this on two fronts:First, build stronger access control and authentication systems, such as "heavy KYC frontier models" for domestic US use and tiered access mechanisms for allies, to reduce the risk of model distillation and misuse. This could also be accompanied by introducing an identity verification system akin to "AI-grade banking authentication" (e.g., biometrics + short-lived permission tokens) to raise the barrier for model calls, and using regulatory measures to restrict account sharing and access resale.Second, enhance the cost efficiency of inference models, allowing them to comprehensively outperform competitors like DeepSeek in both price and performance.Serenity also noted that some high-end models are currently frequently targeted for "distillation exploitation." Ideally, access to models nearing the AGI level should involve increased friction costs. In summary, the core challenge for the US AI industry lies in achieving both "low-cost inference capabilities" and establishing model access security mechanisms comparable to those in the financial system.
Sources: NVIDIA plans to pitch Vera AI CPU to Chinese clients, some cloud providers eyeing test deployment
sources say NVIDIA has begun pitching its first independent central processing unit (CPU) product, Vera, to Chinese clients. Designed specifically for Agentic AI systems, the chip has entered mass production, marking NVIDIA's attempt to further expand its presence in the Chinese market with a CPU offering.According to sources, some Chinese clients have already shown interest in Vera. One major Chinese cloud computing company plans to procure over 300 servers equipped with dual Vera CPUs for testing, and will decide whether to expand procurement after the tests are completed.Built on the Arm Holdings architecture, Vera is NVIDIA's first independent CPU product. NVIDIA has previously stated that Vera's performance in AI agent-related computing tasks is 1.8 times that of comparable competitor products, and expects the product to contribute approximately $20 billion in revenue by the end of this fiscal year (ending January next year).The report notes that as the AI industry's focus gradually shifts from model training to inference computing, CPUs and custom chips are gaining more attention. Vera also positions NVIDIA to directly compete with Intel and Advanced Micro Devices (AMD), which have long dominated the server CPU market.Sources indicate that due to strict U.S. export restrictions on high-end GPUs, CPUs face relatively smaller regulatory hurdles in the Chinese market compared to GPU products. Currently, some Chinese clients plan to first deploy Vera chips for testing in overseas data centers. Meanwhile, software ecosystem compatibility and existing domestic AI chip deployment frameworks may still impact the subsequent large-scale adoption of Vera. (Reuters)
Barclays: Hot Chips 2026 Signals Shift in AI Race From Peak Compute Power to Energy Efficiency Optimization
According to Chaoxiang research, a Barclays report released on August 31 notes that the core theme of the Hot Chips 2026 conference has shifted from "who has the strongest compute" to "who can deliver the most output with the least power," making token/watt the most critical metric for measuring AI hardware success. Constrained by data center power supply, hardware vendors and hyperscalers alike are pushing system performance to its limits within fixed power budgets. Inference workload partitioning has emerged as a key architectural divergence: NVIDIA and Cerebras employ a partitioned approach that decouples prefill and decode stages, delivering roughly an order-of-magnitude improvement in the token/kW metric, whereas OpenAI's Jalapeno integrates prefill, speculative generation, and decoding entirely on a single chip. Barclays anticipates that both approaches will coexist long-term, with the partitioned strategy likely capturing a larger share over time.
AMD middle managers accused of diverting multiple engineers to ATOM inference engine project with almost no customers
Odaily News: SemiAnalysis posted on X that AMD middle managers are spending the time of more than 20 engineers trying to expand their sphere of influence and secure promotions through a useless side project called the "ATOM Inference Engine," rather than allocating those resources to inference engines actually being used by real customers like Meta and xAI, such as SGLang and vLLM.SemiAnalysis said that dozens of AMD engineers, including principal engineers and AMD Fellows, have privately contacted the firm, stating that due to internal office politics at AMD and support from some middle managers, they are unable to discuss their concerns internally. Middle managers are consuming the time of more than 20 full-time engineers to develop this side project with almost no customers, and the engineers involved are all "10x engineers" who should have been dedicated to developing customer-facing inference engines like vLLM and SGLang.SemiAnalysis said that middle managers will likely respond by pointing to the only customer using the project in a production environment—Alibaba, based on Qwen—but further investigation shows that this is merely a limited deployment of ATOM by Alibaba's enterprise business unit, not the main business unit behind Qwen.SemiAnalysis stated that AMD management has diverted resources and limited internal GPU R&D clusters away from engineers supporting production-grade inference engines such as vLLM and SGLang, instead using them to expand their own influence and engage in office politics, showing a profound lack of respect for AMD engineers.
Anthropic Upgrades Claude with Budget Controls, Regional Inference, Skill Loading, and Model Advisor Features
Odaily News – Anthropic has upgraded Claude this week, introducing four significant updates to Claude Managed Agents to further enhance the controllability, deployment flexibility, and task execution capabilities of enterprise-grade AI agents, including:1. New session budget control feature: Users can now set budget limits for agent sessions to achieve more precise cost management. When a session reaches its budget limit, the system will trigger an event and pause operations; users can increase the budget to resume task execution.2. Regional inference control capability: Users can now select where Claude Managed Agents run, with scheduling across globally available resources billed at standard rates. If restricted to run within US regions to meet regional deployment requirements, fees are 1.1x the standard price.3. Support for automatically loading Skills modules from user repositories;4. New model advisor feature: Users can configure a more powerful model as an "advisor," allowing task-executing agents to call upon the advisor model for second-opinion analysis during sessions, improving the quality of complex task handling.
AMD launches AI programming platform Instinct Coder, which can reduce enterprise AI coding costs by 70%
Odaily News AMD, the semiconductor giant, announced the launch of its enterprise-grade AI programming platform, AMD Instinct Coder. The platform combines AMD chips, Supermicro servers, and Spectro Cloud software, aiming to help enterprises deploy AI coding assistants locally, reduce the cost of cloud-based AI models, and protect code and data security.AMD stated that Instinct Coder is an "out-of-the-box" end-to-end AI development platform, integrating AMD EPYC processors, AMD Instinct GPUs, Supermicro AI servers, Spectro Cloud PaletteAI Inference Launchpad software, and the AMD-optimized GLM-5.2 model. It can be used for software development scenarios such as code generation, application modernization, automated testing, and code review.AMD said that compared to relying on cutting-edge cloud-based AI models, Instinct Coder can help enterprises reduce total cost of ownership (TCO) by up to 70%, with the fastest payback period shortened to 6 months.AMD noted that more and more enterprises are looking to leverage AI to improve development efficiency, but face two major challenges: on one hand, the cost of invoking top-tier cloud models continues to rise; on the other hand, entrusting enterprise source code, intellectual property, and sensitive data to third-party services poses security and compliance risks.Through a local deployment model, Instinct Coder allows enterprises to maintain control over their data and code while providing more predictable infrastructure costs. The platform supports development tools such as Claude Code, OpenAI Codex, Visual Studio Code, and Cursor, with each node supporting up to 50 users (30 concurrent users).Additionally, the PaletteAI Inference Launchpad provided by Spectro Cloud enables AI workload management, model routing, request auditing, and cost monitoring, and supports invoking external models such as Anthropic, OpenAI, Google, or xAI when needed.AMD stated that Instinct Coder aims to help enterprises break free from the high costs of cloud-based AI services, accelerate AI-driven software development processes while ensuring data security and autonomous control.
AI Reasoning Startup General Compute Secures $400 Million Loan Backed by ASIC Chips for Inference
: AI reasoning cloud startup General Compute has obtained a $400 million loan from Upper90. This deal is the world’s first financing project to use dedicated inference chips as collateral. The company has built a proprietary AI reasoning cloud platform based on SambaNova’s self-developed ASIC chips, primarily targeting Agent-type AI computing workloads. Compared to traditional GPU clouds, it offers faster token processing speeds and lower operational latency. The hardware requires no water cooling and can be directly deployed in traditional data centers and idle cryptocurrency mining facilities.
Meta's self-developed AI chip "Iris" is planned to start mass production in September, with a 2027 compute target of 14 gigawatts.
According to Reuters, Meta plans to mass-produce its self-developed data center AI chip "Iris" starting from September, as part of its fourth-generation Meta Training and Inference Accelerators project, to enhance the AI capabilities of platforms such as Facebook and Instagram and reduce reliance on external GPUs such as those from Nvidia and AMD. Internal memos show that Iris completed testing in just 6 weeks with no major defects; Meta plans to deploy 7 gigawatts of computing power this year and increase it to 14 gigawatts by 2027, with its AI infrastructure spending in 2024 potentially reaching up to $145 billion. To secure expansion, the company has signed long-term supply agreements with Samsung Electronics, Sandisk, and Sumitomo Electric to cope with "price increases" and shortages of memory and AI chips.
Related news
Morgan Stanley: AI Inference Era Begins, CapEx to Slow in 2028, Funds Shift to Software Layer
According to TrendForward research, a Morgan Stanley research report dated September 4, 2026, indicates that hyperscaler data center capital expenditures will reach $1.5 trillion in 2027, a 60% year-over-year increase, before the growth rate drops sharply to 12% in 2028. Total AI compute capacity is expected to expand from 35 gigawatts in 2025 to 145 gigawatts in 2028, representing a fourfold increase. Custom chips will capture a growing share of incremental capacity, rising from 34% to 66%, driven primarily by Google TPUs and Amazon Trainium. GenAI returns on investment range between 25% and 50%, with the model layer delivering the highest returns by running APIs on proprietary infrastructure (approximately 46%). Approximately 25% of S&P 500 companies have quantified GenAI returns, up from 14% a year ago.
Morgan Stanley: 70% of NVIDIA's FY28 Revenue Guidance Constrained by Supply; Inference Has Become the Largest and Continuously Expanding Data Center Segment
According to Chaohiang Research, JPMorgan's September 2 investor meeting minutes for NVIDIA indicate that NVIDIA is "comfortable" with its guidance for 70% year-over-year growth in FY28, and this guidance is supply-constrained rather than demand-constrained; with sufficient supply, the business could grow more than double. Inference has become the largest and expanding segment of the data center business. Eighteen months ago, inference and training each accounted for roughly half, but currently inference exceeds training and will continue to rise. Advanced wafers and memory are the two major supply bottlenecks, and NVIDIA maintains close cooperation with TSMC and the three major memory suppliers. The customer base continues to broaden, with OpenAI and Anthropic currently accounting for approximately 20% of end-user demand, which may approach 25% by FY28; new cloud providers now account for over 50% of AI computing infrastructure. Open-source and closed-source models will coexist, and gross margins for model developers are improving. NVIDIA supports long-term demand through revenue sharing, the PORTS-Pike campus, and a $500 billion private capital financing platform. Morgan Stanley maintains an Overweight rating with a price target of $320, based on approximately 20x the expected CY2026 EPS of $15.87.
Barclays: Hot Chips 2026 Signals Shift in AI Race From Peak Compute Power to Energy Efficiency Optimization
According to Chaoxiang research, a Barclays report released on August 31 notes that the core theme of the Hot Chips 2026 conference has shifted from "who has the strongest compute" to "who can deliver the most output with the least power," making token/watt the most critical metric for measuring AI hardware success. Constrained by data center power supply, hardware vendors and hyperscalers alike are pushing system performance to its limits within fixed power budgets. Inference workload partitioning has emerged as a key architectural divergence: NVIDIA and Cerebras employ a partitioned approach that decouples prefill and decode stages, delivering roughly an order-of-magnitude improvement in the token/kW metric, whereas OpenAI's Jalapeno integrates prefill, speculative generation, and decoding entirely on a single chip. Barclays anticipates that both approaches will coexist long-term, with the partitioned strategy likely capturing a larger share over time.
AMD middle managers accused of diverting multiple engineers to ATOM inference engine project with almost no customers
Odaily News: SemiAnalysis posted on X that AMD middle managers are spending the time of more than 20 engineers trying to expand their sphere of influence and secure promotions through a useless side project called the "ATOM Inference Engine," rather than allocating those resources to inference engines actually being used by real customers like Meta and xAI, such as SGLang and vLLM.SemiAnalysis said that dozens of AMD engineers, including principal engineers and AMD Fellows, have privately contacted the firm, stating that due to internal office politics at AMD and support from some middle managers, they are unable to discuss their concerns internally. Middle managers are consuming the time of more than 20 full-time engineers to develop this side project with almost no customers, and the engineers involved are all "10x engineers" who should have been dedicated to developing customer-facing inference engines like vLLM and SGLang.SemiAnalysis said that middle managers will likely respond by pointing to the only customer using the project in a production environment—Alibaba, based on Qwen—but further investigation shows that this is merely a limited deployment of ATOM by Alibaba's enterprise business unit, not the main business unit behind Qwen.SemiAnalysis stated that AMD management has diverted resources and limited internal GPU R&D clusters away from engineers supporting production-grade inference engines such as vLLM and SGLang, instead using them to expand their own influence and engage in office politics, showing a profound lack of respect for AMD engineers.
OpenAI Releases Test Results for First Custom Inference Chip Jalapeño, Leading in Energy Efficiency and Latency
OpenAI announced the preliminary test results for its first custom inference chip, Jalapeño. Data indicates that compared to baseline systems, the chip delivers a 1.5x to 1.9x increase in peak performance per watt across models including GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, while reducing end-to-end latency by 1.7x to 3.6x; under high-interaction workloads, performance improvements reach 2.1x to 4.1x.
Morgan Stanley: CoreWeave Adds Record 500MW Capacity in Single Quarter, High Debt and Customer Concentration Weigh on Valuation
According to TechFlow Research, Morgan Stanley's August Q2 earnings report indicated that CoreWeave added 500MW of net active power in a single quarter, exceeding any quarter in history, more than three times year-over-year. Management reaffirmed the target of reaching at least 8GW before 2030. FY26 revenue guidance midpoint was raised 2% to $12.4 billion to $13.2 billion, ARR midpoint was raised 3% to $18.5 billion to $19.5 billion. The company raised full-year capital expenditure guidance midpoint by 12% to $35.5 billion to $39.0 billion, Q3 capital expenditure guidance is $11.5 billion to $13.5 billion, higher than the market expectation of $10 billion. Managed Inference Platform (managed inference platform) ARR grew from $1 million to over $100 million, expected to reach at least $250 million by year-end. The research report judges that Q2 adjusted operating margin was about 8%, higher than expected, but Q3 margin guidance of 5.8% to 7.2% is lower than market expectations, Q4 margin needs to increase significantly to achieve full-year guidance. Morgan Stanley expects CoreWeave FY27 operating margin to be 15.9%, FY28 to be 22.4%, free cash flow to remain negative until 2028, and debt is expected to increase to approximately $38 billion by the end of 2026. Morgan Stanley maintains Equal-weight (in line with the market) rating and $99 price target.