News linked to both this project and an event.
Odaily News, Citrini analyst jukan stated on the X platform that at first glance, SK Hynix's CPO technology roadmap appears to have one axis pointing toward HBM and another toward optical communication. However, the underlying logic is that AI competition is shifting from individual chip performance to data transfer efficiency across the entire system.Over the past few years, HBM has addressed the problem of GPUs being unable to receive data fast enough. By vertically stacking multiple layers of DRAM and placing them next to the GPU, HBM delivers extremely high bandwidth. Yet, as more HBM stacks are placed around each GPU, packaging area, interposer edge space, power supply, and thermal dissipation capabilities are all approaching their limits. Meanwhile, AI clusters have expanded to thousands or even tens of thousands of GPUs. No matter how fast a single GPU computes, the entire system will still be constrained by the "bandwidth wall" if data cannot be efficiently transferred between GPUs and racks.SK Hynix's vision is to extend optical interconnect to memory. Low-latency, high-bandwidth local HBM will remain next to the GPU, while optical fibers connect it to a larger shared memory pool. This approach avoids limiting all memory capacity within a single GPU package and enables horizontal scaling at the rack level.From an investment perspective, this does not mean HBM will be replaced in the near term. What is more likely to emerge is a new memory hierarchy: frequently accessed data remains stored in local HBM, while larger-scale, less frequently accessed data is stored in optically interconnected memory pools, HBF, or SSDs.SK Hynix is repositioning its business, shifting from selling standardized memory chips to co-designing HBM, controllers, advanced packaging, and system-level memory architectures with customers. If this roadmap comes to fruition, SK Hynix could strengthen customer stickiness, increase product added value, and enhance its ability to secure long-term contracts. At the same time, the company may also more proactively address the potential impact of future memory disaggregation on traditional HBM business models.At the supply chain level, areas that may benefit in the long term include silicon photonics chips, optical engines, lasers, fiber coupling technology, and advanced 2.5D and 3D packaging.However, this concept is still in its very early stages and essentially remains just a roadmap. jukan noted that a person involved in TSMC's packaging business whom he interviewed today was completely unaware of this plan.
Odaily News, Citrini analyst Zephyr posted on X platform questioning that SanDisk's HBF vs. HBM performance comparison at its Investor Day may contain parameter configuration bias, arguing that the company deliberately underestimated HBM performance in its demonstration.Zephyr pointed out that SanDisk set total bandwidth for both HBM and HBF at 12.8TB/s, while running the Qwen3-480B-A35B model with bfloat16. Current model inference more commonly uses FP4 or FP8, requiring approximately 240GB to 480GB of capacity.Zephyr also noted that SanDisk fixed HBM capacity at 192GB per GPU, but 16-layer HBM4E in an 8-stack configuration can reach 512GB capacity with approximately 32TB/s bandwidth—roughly three times the parameters used in SanDisk's demonstration. Zephyr believes these parameter choices result in the company's claim that "HBF requires only 1 to 4 GPUs, while HBM requires 8 GPUs" failing to adequately represent actual performance under higher-spec HBM configurations.
Odaily News - Analyst qinbafrank posted on X platform, stating that the latest earnings reports from CoreWeave (CRWV) and Nebius show the AI cloud computing (CSP) industry is entering a phase of rapid expansion. The competitive focus is shifting from simply providing GPU leasing to building AI infrastructure platforms that encompass computing power, software, data, and operational capabilities.Currently, AI computing demand still significantly exceeds short-term deliverable supply. Meanwhile, pricing power for AI computing is strengthening, but price increases are mainly concentrated on high-value resources. CoreWeave stated that prices for various GPU computing SKUs rose by approximately 25% on average in July; Nebius disclosed that prices for previous-generation GPUs increased by over 30% compared to Q1, with new contracts signed in Q2 averaging over $20 million in annualized revenue per MW, some projects reaching $20 million to $25 million, and short-term emergency capacity prices even reaching $40 million to $50 million per MW.However, price increases are mainly occurring in short-term capacity, next-generation GPUs, large-scale clusters, and production-grade AI inference scenarios. Traditional low-priority, long-term locked-in bare computing power has not seen concurrent price increases. From a profitability model perspective, project-level returns on AI computing are becoming clearer, but overall corporate return on invested capital (ROIC) still needs time to be validated. Nebius has for the first time disclosed relatively clear project payback periods, while CoreWeave is reducing GPU investment pressure through long-term contracts and asset-level financing. However, both companies remain in a high-capital-expenditure phase, with depreciation and financing costs continuing to compress profit margins.Nevertheless, an increasing number of individual projects are achieving closed-loop economic models, indicating that the AI infrastructure business model is gradually maturing. Additionally, both CoreWeave and Nebius are upgrading toward becoming "AI infrastructure operating systems." Future CSP competition will no longer be just about renting out GPU hours but will cover complete service systems including AI training, inference, storage, networking, model deployment, monitoring, security governance, and Agent runtime environments.In terms of capital models, the two companies are also taking different paths: Nebius leans more toward an asset-light model, building AI data centers through capital partners while providing AI infrastructure operations and software capabilities itself; CoreWeave, on the other hand, is promoting a hybrid cloud model through its Omni strategy, deploying complete AI cloud platforms to customers' own data centers and GPU resources, placing greater emphasis on enterprise-level and sovereign AI delivery.Overall, the AI cloud computing industry is evolving from "GPU rental providers" to "AI infrastructure platforms." Short-
Odaily News, Citrini analyst jukan posted on X platform, according to TrendForce's latest memory industry research, the DRAM supply shortage is expected to persist through 2027, and the HBM4e certification timeline remains uncertain. NVIDIA has been reassessing the HBM configuration for Rubin Ultra since the third quarter of 2026.The original plan called for a 12-layer stacked HBM4e configuration, but NVIDIA is currently evaluating multiple designs in parallel, including HBM4e 8-layer stacking, HBM4 12-layer stacking, and HBM4 8-layer stacking, with the final specifications yet to be confirmed. In addition to NVIDIA, some CSPs are reportedly also considering reducing HBM capacity for their next-generation custom ASICs.From 2025 to the first half of 2026, NVIDIA used 12-layer stacked HBM4e as the baseline design for Rubin Ultra. However, since the beginning of the third quarter of 2026, NVIDIA has begun reviewing lower-spec alternatives. TrendForce attributes this change to two major supply-side constraints: first, the overall DRAM shortage expected in 2027 will limit the wafer capacity that memory manufacturers can allocate to HBM production; second, uncertainty remains regarding the certification timeline for 12-layer stacked HBM4e and the pace of yield improvement in mass production.TrendForce stated that NVIDIA's primary focus for the Rubin Ultra generation is improving I/O speed, with expanding GPU shipments as a secondary priority. If NVIDIA ultimately decides to downgrade the HBM specifications, it is expected to do so by reducing the number of DRAM stacking layers. Whether HBM4e can complete certification and enter mass production as planned will determine whether Rubin Ultra's I/O speed can be improved from the previous generation Rubin's 8 to 11.7Gbps to 14 to 16Gbps, or be maintained at 11 to 12Gbps through an optimized HBM4 design. Within the same product generation, the number of DRAM stacking layers determines the trade-off between HBM capacity per GPU and the number of GPUs that can be shipped.TrendForce also believes that the final configuration will depend on wafer allocation decisions made by memory manufacturers. On the supply-demand front, HBM bit shipments in 2027 are expected to grow 50% to 60% year-over-year, but are still projected to fall short of demand growth. With supply constraints persisting, HBM suppliers are expected to maintain pricing power throughout 2027. The industry has broadly anticipated significant HBM price increases, and AI chip makers will face the dual pressure of limited HBM supply and rising procurement costs, further strengthening the incentive to adopt lower HBM capacity configurations.
analyst KawzInvests stated that Moonshot AI's upcoming Kimi K3 could become a significant event in the open-source AI space, and the infrastructure demand behind it may drive growth for AI cloud service platforms. Kimi K3 has approximately 2.8 trillion parameters, making it one of the ultra-large-scale open-source models. According to Moonshot's official evaluation, the model's performance is only slightly behind frontier models like Claude Fable 5 and GPT 5.6 Sol, and the full model weights are expected to be released on July 27.KawzInvests pointed out that a model of this scale cannot run on an ordinary laptop or even a single server; users need a computing cluster composed of a large number of GPUs to complete model loading and inference. When top-tier open-source models are made available for free, the biggest beneficiaries might not be ordinary users, but rather platforms that offer model hosting and inference services. For example, $DOCN (DigitalOcean) already supports serverless inference services for models like Kimi K2.6. Developers do not need to deploy hardware; they can call the model via API and pay per Token. Additionally, the platform hosts over 70 models and covers GPU leasing, model fine-tuning, and AI Agent development tools.As more large-scale open-source models are released, developers' demand for low-barrier AI infrastructure will continue to increase. Model hosting, inference services, and GPU cloud platforms may become key beneficiaries in the open-source AI wave.
According to BitcoinTreasuries data, South Korean listed company K Wave Media (KWM) has sold all of its remaining 88 BTC to repay $6 million in debt. Following the sale, the company's Bitcoin holdings have dropped to zero, exiting the ranks of Bitcoin treasury companies.K Wave Media announced last year that it had secured a $1 billion Bitcoin treasury financing capacity and planned to expand its Bitcoin holdings to 10,000 BTC as soon as possible. However, in May this year, the company redirected up to $485 million of its remaining financing capacity from the Bitcoin treasury strategy to AI infrastructure construction, including data centers, GPU computing power, and related acquisitions.
: In response to the community debate over "Micron VS Nvidia," Jukan, an analyst at Citrini Research, posted on platform X, stating, "MU (Micron) may not be Nvidia, but its future importance could surpass Nvidia. Inference is now directly tied to revenue, but improvements in inference performance cannot be achieved simply by adding more Nvidia GPUs. In inference, GPUs are often underutilized and idle due to memory bottlenecks. For inference, increasing memory provides higher value. The return on investment for inference ultimately depends on memory, not GPU. Therefore, why are people still fixated on obtaining Micron through Nvidia's framework? One must think more comprehensively. Inference is memory."
Odaily News Alex Svanevik, CEO of on-chain analytics platform Nansen, stated that when enterprises begin effectively using Chinese large language models, the bubble in the AI industry may burst. While the U.S. regulatory environment could limit this process, the overall trend remains that Chinese models are continuously becoming more efficient, capable of running on non-cutting-edge hardware, while global GPU supply (including non-Nvidia chips) is increasing.Alex Svanevik also pointed out that the recent decline in H100 and H200 GPU rental prices reflects a shift in the supply-demand structure of computing power. He raised the question of how to interpret the market signal of declining GPU rental prices. As model efficiency improves alongside expanding computing power supply, the AI infrastructure market may be entering a phase of repricing.
sources familiar with the matter have revealed that Goldman Sachs and JPMorgan are exploring trading methods based on the cost of computing power, including futures contracts linked to GPU rental prices. As one of the scarcest resources amid the AI boom, related futures for GPUs are expected to be listed on exchanges later this year.Industry insiders stated that this move reflects how the influx of hundreds of billions of dollars into data centers and the chip sector is reshaping the financial market landscape. For banks financing the construction of AI infrastructure, such innovative instruments could become a new means of risk management. (The Information)
"New Stock God" Serenity posted on X platform, reminding investors to pay attention to financing structures and the dynamics of outstanding shares, as these are crucial for investment returns, and provided examples:IREN: The financing method approaches infinite dilution, with each rebound met by selling pressure—essentially a "bad stock."NBIS: Up 153% year-to-date, thanks to an optimized financing structure (such as direct offerings, convertible bond combinations, etc.).CRWV: High debt interest; the company uses usurious loans for GPU financing, which erodes free cash flow over the long term.Serenity pointed out that if a company has strong fundamentals, one could consider going long after the original shareholding has been diluted to near zero. However, for equity value appreciation, one should stay away from companies with "toxic" financing structures or crushing debt. The risk is especially high for small-cap companies, such as $SLNH adding a $500 million ATM while its market cap is only $250 million; $BKKT continuously diluting stock for executive compensation. Essentially, these companies are transferring investor funds to the enterprise, masked by media hype or influencer promotion.Serenity emphasized that investors must carefully analyze equity structure, dilution risk, and hidden costs when screening targets, to avoid focusing solely on profits while seeing their actual equity shrink.
According to Citrini analyst Jukan, FundaAI's latest report reveals that NVIDIA's NVL576 passive co-packaged optics technology nearly doubles the density of optical engines and optical components. The passive co-packaged optics technology increases the configuration volume of 3.2T optical engines per GPU from approximately 2.25 to around 4.0, an increase of 78%. It is estimated that the demand for Rubin Ultra optical engines will reach about 12 million units.This news may directly benefit direct suppliers of optical engines, including Lumentum, Coherent, POET, among others.
AethirClaw has officially launched CARA (Pre-configured Crypto AI Agent), running on Aethir's decentralized GPU infrastructure. Equipped with over 50 skills, it covers core crypto scenarios such as real-time market monitoring, whale wallet tracking, on-chain analysis, social media sentiment monitoring, and project due diligence, and users can use it out-of-the-box without any technical configuration.The platform supports payments via credit card as well as USDT, USDC, and ATH tokens. Aethir also disclosed that it will soon launch a Model-as-a-Service (MaaS) layer, running mainstream open-source large models on Aethir's decentralized GPU infrastructure, and expand multimodal capabilities including text-to-image and video generation.