semianalysis

SemiAnalysis released the latest episode of its podcast, which pointed out that SK Hynix and Micron have been facing challenges in achieving the highest HBM4 speeds. In particular, SK Hynix had to redesign its base chip, resulting in delays in HBM4 deliveries to the NVIDIA Rubin platform.Samsung currently has the best HBM4 technology. Broadcom has traditionally relied heavily on Samsung for its HBM supply. Samsung's performance in the HBM3E generation was poor, which posed a disadvantage for Broadcom's ASICs. However, now Samsung is leading its competitors with HBM4, and this same supply chain relationship has become an advantage for Jalapeño.

The semiconductor and AI independent research organization SemiAnalysis released a deep security report on Neocloud (new cloud), revealing various cross-tenant security vulnerabilities discovered during the ClusterMAX 3 testing period. In a four-month test covering 25 vendors and 32 clusters, the team achieved multiple instances of cross-tenant remote code execution (RCE) solely by exploiting publicly known vulnerabilities and basic configuration checks. Affected entities included banks, telecommunications companies, universities, research institutions, AI laboratories, and even a national intelligence agency.Typical issues included: shared Kubernetes control plane leading to tenant metadata visibility, container escape, exposure of BMC/IPMI management networks, incorrect configuration of InfiniBand security keys (P_Key, SA_Key, M_Key), unfortified default trust mode of BlueField DPU, Grafana monitoring dashboards using god-level API keys, and lack of VXLAN isolation in front-end networks. The report specifically pointed out a cascading vulnerability case: a misconfiguration of shared vCluster combined with software versions being two years out of date ultimately completed the POC verification of cross-tenant RCE within an afternoon.Notably, the report questioned the mainstream narrative that "AI has fundamentally changed the pace of cybersecurity": statistics on CVEs for NVIDIA GPU drivers, CUDA, PyTorch, Kubernetes, Docker, and the Linux kernel showed that there was no significant increase in vulnerabilities after the popularization of AI coding models, with most data supporting the "no change hypothesis." The report also detailed the incident where an OpenAI-trained agent attacked Hugging Face, where the AI agent achieved cluster-level privilege escalation through a message board established via Artifactory, which went undetected from May to July. While building POC verification for existing vulnerabilities, the team found that Claude Fable and GPT-5.6 Sol frequently rejected security-related requests, ultimately relying on open-source models such as DeepSeek V4, Kimi K3, and GLM-5.2 to complete the task.SemiAnalysis stated that the core issue in the Neocloud (new cloud) industry is not the new risks brought by AI, but rather the long-term absence of basic patch management, tenant isolation, and security design. They recommended that vendors establish automated security announcement monitoring systems and rectify single points of failure that could expose all users' architectural patterns.

P Equity Research and SemiAnalysis researcher Nick Doyle and others discussed high bandwidth flash (HBF) in X Space. Nick stated that it is still too early to determine how much the cost premium of HBF relative to HBM can shrink; existing data mostly comes from vendor claims, such as Sandisk stating that the cost per bit is about one-eighth that of HBM. Yields, testing, and other factors will improve with scale, but structural costs such as TSV, stacking, and pSLC mode will always exist, and durability is a key unknown; if wear exceeds expectations, costs will rise.The application scenarios for HBF are narrow, targeting only AI inference, especially low batch and long context MoE models, and it is not a substitute for HBM. The actual bandwidth target is about 1.6 TB/s, which is at the HBM3E level, suitable for sequential reads to load model weights, more aligned with the capacity needs of a small number of GPUs in local or private enterprises, rather than ultra-large-scale bandwidth scenarios. Heat dissipation reliability has not yet been resolved; flash memory will degrade faster at high temperatures next to GPUs, and mitigation measures such as UCIe separation and daily refresh have yet to be validated.In terms of manufacturing, Sandisk/Kioxia has experience with 3D NAND, and SK Hynix complements HBM-style stacking capabilities, but mass production is still to be confirmed. Overall, storage is shifting towards a specialized layered market, with NAND shortages expected to continue until 2028, and HBF may further impact supply and demand.

In the latest podcast, SemiAnalysis founder Dylan Patel predicts that by 2028, OpenAI and Anthropic may account for 70% to 80% of the world's new AI computing power, with the total computing power scale potentially exceeding 100GW. Patel stated that the two companies currently account for about 30% of the world's annual new computing power, and this proportion is still rising rapidly.Patel pointed out that the business model of leading AI laboratories is changing, with a significant increase in the efficiency of AI computing power output. Currently, Anthropic's revenue per megawatt of computing power has reached about $50 million and may further rise to $100 million. This allows OpenAI and Anthropic to procure or lease computing power at high prices ranging from $25 million to $50 million per megawatt.Patel expects that global AI-related capital expenditures will reach about $11 trillion from 2024 to 2029, with over $5 trillion needing to be financed through debt. Due to the potential return on investment of AI infrastructure being far higher than that of traditional industries, tech giants may accept higher financing costs, thereby pushing up overall credit rates and squeezing the valuations of traditional assets and highly leveraged economies. Additionally, Patel believes that the new computing power may not primarily be used for providing model inference services externally, but may instead flow more towards internal research and development and self-improvement of models within AI laboratories.

SemiAnalysis reported that the Nvidia Rubin Ultra will feature 192GB of HBM4 8-hi memory. This is not only a significant reduction from the initially previewed 1TB HBM but is even lower than the 288GB capacity of the standard Rubin.The reasons for the HBM configuration being cut down to about one-fifth of the original include: the cancellation of the four-die Rubin Ultra scheme, reducing the number of cubes to 8; the stacking height being lowered from 16-hi to 8-hi; and the use of HBM4 with 24Gb dies instead of HBM4E with 32Gb dies. An upgraded version with 8-hi HBM4E may be released later.

SemiAnalysis published that it tested Anthropic's $200 per month Claude plan, which can produce tokens worth up to approximately $8,000 per month if the relevant limits are exhausted in long-term tasks; in contrast, OpenAI's $200 per month plan can reach a token value of about $14,000 per month when all weekly limits are exhausted.The test was completed in June of this year. SemiAnalysis pointed out that the actual usable model value corresponding to the high-priced subscription plans of both companies in high-intensity usage scenarios far exceeds the subscription fees themselves.

According to SemiAnalysis, Anthropic's ARR from indirect channels (Bedrock, Foundry, Gemini Agent Enterprise) in Q2 2026 has exceeded 40%, with API and B2B contributing the majority of new revenue. The monetization model for indirect channels differs from direct revenue; cloud service providers typically charge IaaS fees or revenue sharing. Laboratories account for gross ARR, but related costs are reflected in sales and marketing expenses. Indirect channels require precise computational resource allocation analysis at the individual transaction level.

According to the SemiAnalysis podcast, Anthropic has set extremely strict classifiers for the public version of its flagship model Mythos, called Fable, due to regulatory pressure. Users attempting specific actions will be immediately downgraded to the Opus model, and conversations cannot be reverted. Dylan Patel, founder of SemiAnalysis, stated that the classifier restrictions are "too harsh," and even internal employees cannot complete routine tasks with it.The podcast mentioned that the Mythos model completed training as early as February, but Anthropic did not release it under the name Fable until May. Jordan stated in the discussion that, to his knowledge, Mythos 2 has completed training but has not been released, and the internal R&D team is believed to still be iterating. Analysts believe that Anthropic's previous long-standing calls for regulation have now resulted in regulatory measures that limit the release pace of its own models. Similarly, OpenAI is also delaying the public release of its next-generation models, but internal training is still progressing, and the capability gap between internal models and public models may continue to widen.

Research institution SemiAnalysis released an analysis stating that SpaceX is expected to build approximately 10GW of AI data center capacity by the end of 2027. If 50% of this is used for inference services, with annual revenue exceeding $10 billion per GW, the annualized revenue could reach $300 billion. SpaceX CEO Elon Musk stated in the first earnings report that a "conservative estimate" suggests an additional 6-8GW will be added in 2027, with the actual figure possibly exceeding 10GW.SemiAnalysis's inference simulator shows that when running on the GB300 cluster at current startup cloud prices (about $3/GPU hour), leading model companies like OpenAI and Anthropic could generate annual inference revenue exceeding $10 billion per GW, with annual costs around $12 billion. Microsoft, with full access to OpenAI models and without bearing training costs, can also capture revenue of the same scale. The analysis points out that Microsoft has signed contracts for 10GW of data centers (total value exceeding $300 billion) since 2026, with a 90-day cancellation clause, significantly reducing signing risks.Regarding SpaceX's construction progress, SemiAnalysis believes Musk will significantly shorten the construction cycle by using onsite gas power generation, bypassing large power transformers, parallel construction, and shortening the debugging process. The Southaven plant in Tennessee expanded from 27 turbines (approximately 495MW) in February to 69 turbines (1.7GW) in July, and the "MiniHard" project can be completed in about 5 months with 450-500MW. However, the 10GW target still faces multiple challenges such as land approvals, gas supply, and equipment delivery. This analysis is based on model simulations, and actual implementation still carries uncertainties.

The research organization SemiAnalysis released an analysis indicating that Google DeepMind is no longer among the leading AI laboratories. A week prior, DeepMind co-founder Demis Hassabis stepped back from daily operations, and key members such as Google Chief Scientist Jeff Dean and Gemini co-lead Oriol Vinyals left to establish a new lab called Discovery Loop. The analysis suggests that the long-term struggle within Google over computing power allocation between Gemini and GCP has concluded with GCP emerging victorious.SemiAnalysis stated that Gemini 3.5 Pro has been canceled, and Gemini 3.6 Flash's performance is inferior to that of leading Chinese open-source models and Grok 4.5. Currently, Gemini has fallen to the 8th or 9th position in the large model rankings. Meanwhile, GCP is selling a large number of TPUs to competitors like Anthropic, having secured long-term leasing and sales contracts for hundreds of thousands of TPUs over the past nine months. The Tokenomics model estimates that Gemini's own ARR is about $12 billion, while GCP's third-party AI cloud service revenue is expected to exceed $73 billion by the end of 2027, with TPU system sales contributing an additional over $120 billion. GCP's latest quarterly growth rate is 82%, and it is expected to accelerate to over 100% by 2027 due to TPU system sales, contributing approximately $3 to Google's earnings per share.

According to SemiAnalysis, AMD reached a performance-based warrant agreement with Meta and OpenAI from the end of last year to the beginning of this year. The warrants will vest in batches based on the two companies' procurement and deployment of up to 6GW of AMD Instinct GPUs, while AMD's stock price must exceed a tiered target (up to $600).As AMD's stock price has risen from around $200 in early 2026 to the current $500-600, SemiAnalysis estimates that the GPU capital costs for Meta and OpenAI have decreased from about $2.15/hour and $2.50/hour to approximately $1.60/hour and $1.75/hour, respectively. If AMD's stock price rises to $600-700, the capital cost of the GPUs themselves will approach zero; if it further exceeds $1000, the total capital cost of the entire cluster (including servers, networks, and CPUs) will effectively be zero for both companies. This analysis is based on the assumption of linear vesting and the premise that the entire 6GW capacity is deployed, and the specific terms of the warrants have not been disclosed.

The semiconductor research organization SemiAnalysis published an article stating that Nvidia is expected to experience a significant ramp-up in production capacity in the second half of this year (i.e., the second half of fiscal year 2027), with its data center computing business revenue projected to exceed current market consensus expectations by 20%.The organization noted that the issues previously delaying Nvidia's next-generation Rubin architecture chips due to technical bottlenecks with HBM4 have now been resolved, and initial reserves of front-end wafer supply have been completed, with large-scale mass production about to begin.In addition, SemiAnalysis emphasized that its forecast data is based on in-depth research across the entire industry chain, covering dynamics from upstream materials, wafer foundries, components, and server integrators, to downstream hyperscale cloud service providers and cutting-edge AI laboratories. Its tracking model also includes major AI chip ecosystem participants such as Broadcom, AMD, MediaTek, and Marvell.

The semiconductor research institution SemiAnalysis has released a latest analysis indicating that Changxin Memory Technologies (CXMT) has clearly become the world's fourth largest DRAM manufacturer. Although its production capacity and cash flow are continuously growing, the institution believes that Changxin Memory still faces multiple challenges in equipment, technology, and market, and will not end the current storage "super cycle" in the short term.In terms of specific challenges, export controls on advanced semiconductor manufacturing equipment (such as EUV, advanced etching, and TSV tools) severely restrict Changxin's expansion into more advanced processes and high bandwidth memory (HBM) fields; although domestic equipment (such as Zhongwei Company, Northern Huachuang, etc.) has alleviated some pressure, it cannot fully resolve the integration and yield bottlenecks across multiple process links, resulting in its technology still lagging behind leading manufacturers by several generations. Additionally, Changxin's market share is currently still highly concentrated in the Chinese domestic market, with global expansion limited by geopolitical factors and customers' willingness to diversify their supply chains.In response to market concerns that Changxin might "impact the global market with cheap chips," SemiAnalysis clarified that there is currently a severe structural shortage in the DRAM market, and the increase in Changxin's production capacity may even struggle to fully meet domestic demand in China. In fact, the prices of Chinese memory chips are also soaring significantly, in line with the global upward trend, and Changxin is similarly a beneficiary of the shortage premium. Therefore, Changxin Memory should be viewed as a long-term structural competitive force, and in the current context of accelerated AI demand and constrained supply, it cannot shake the fundamental super cycle dominated by leading manufacturers in the short term.