DeepSeek’s recent rollout of Harness, released around the same time as the V4 Pro price increase, has prompted a new reading of the company’s strategy. The original MarsBit commentary argues that DeepSeek may be moving away from a playbook centered mainly on cheaper and stronger models, and toward control over where model calls actually happen. Harness is presented not as a polished end-user Agent product, but as a modular runtime framework. In the article’s description, models, tools, Skills, workflows, and UI are all separable parts, while plugins can be loaded, removed, and recombined on the fly. Underneath that sits Cordis, a kernel focused on plugin lifecycle management and system stability during change. The piece links that design to a broader idea: Agents that can adjust their operating environment while they run. Another point the article highlights is that Harness does not force users onto DeepSeek’s own models. Developers can define models, protocols, and Base URL settings, and can theoretically connect competing models. From that angle, the commentary suggests DeepSeek may be targeting the Agent runtime layer as infrastructure, closer to an operating system logic than a single-model product strategy. In that framing, V4 Pro’s higher pricing and Harness’s open design are part of the same shift toward ecosystem economics.

A bill-of-materials report from Aletheia Capital points to a sharp increase in pricing for Nvidia’s upcoming Rubin Ultra AI compute modules. The firm estimates the average selling price of the dual-GPU version at nearly $170,000, while the four-GPU configuration could reach $350,000. That would represent a 30% increase from the previous Rubin generation. The report also shows a major shift in the module’s cost structure: SoCAAM, an advanced packaged memory module, has overtaken the GPU die itself as the largest single cost component and now accounts for more than half of total BOM in the dual-GPU Rubin Ultra version. Analysts cited in the report said Rubin Ultra is expected to move directly to HBM4E instead of HBM4, while a higher-end four-die model remains on the roadmap. They also said the previously rumored-canceled Kyber platform is still moving ahead. If those plans hold, the report suggests memory suppliers, especially leading HBM vendors, could benefit as HBM4E demand scales and supply takes time to ramp.

Sandisk used its Aug. 13 investor day to lay out a long-range financial model that sits well above market expectations and puts its in-house High Bandwidth Flash, or HBF, at the center of the story. The company said its long-term sustainable model, based on average performance across FY2028 to FY2030, calls for revenue growth in the mid-to-high teens, a Non-GAAP gross margin of about 80%, a Non-GAAP operating margin of roughly 75%, adjusted free cash flow margin near 50%, and capital intensity in the mid-single digits as a percentage of revenue. Mizuho Securities kept its Outperform rating on Sandisk and set a $1,900 price target, arguing that the company’s framework, buyback capacity, and exposure to AI infrastructure support upside. Sandisk also detailed the positioning of HBF, saying the product can deliver comparable read bandwidth to HBM at about one-eighth the cost, while offering 8x to 16x more capacity in a similar package footprint. The company tied that pitch to memory-heavy AI inference workloads and said HBF is meant to work alongside HBM rather than replace it.

Ethereum is moving away from Poseidon at Layer 1, ending years of work around a hash function long favored in SNARK-based systems. On Aug. 13, Ethereum researcher Justin Drake said on X that the Ethereum Foundation had decided to abandon Poseidon in favor of traditional hash functions such as SHA2 or BLAKE2. The change comes after eight years of research, tens of millions of dollars in spending, and a broader reset of Ethereum’s post-quantum roadmap. The key technical driver is progress in binary-field SNARK design, which allows traditional hash functions to perform inside proving systems at speeds that were previously associated with SNARK-optimized designs. Drake said a laptop can now verify about 1 million conventional hash calls per second in a SNARK setting, while recent benchmarks from projects including Flock and SNARK.fast point to sharply improved throughput. The roadmap itself remains in place. Ethereum still expects a production-grade leanVM in 2027, followed by deployments across the consensus, execution, and data layers in 2028. The Foundation has also expanded its post-quantum work through pq.ethereum.org, weekly interoperability devnets involving more than 10 client teams, and two $1 million research prizes. The shift also comes as Solana and Starknet advance their own post-quantum plans using Falcon and BLAKE2-based transitions.

Advanced Micro Devices said it plans to issue up to $5 billion in senior unsecured notes, a deal that would rank as the company’s largest U.S. dollar bond sale on record. The offering is split across four maturities ranging from three to 10 years, with final size dependent on investor demand. AMD said the proceeds will be used for general corporate purposes, including debt repayment, as $875 million of its bonds come due next month. As of June 27, 2026, AMD held about $5.09 billion in cash and cash equivalents and $8.03 billion in short-term investments. The transaction is being underwritten by Barclays, BofA Securities, Citigroup, J.P. Morgan, Morgan Stanley, and Wells Fargo. ABMedia said the fundraising is tied to AMD’s broader AI expansion. The company has recently worked with Anthropic and Microsoft, and has committed up to $5 billion to Anthropic to support and optimize Claude model performance on AMD hardware. AMD is also targeting more than 50% share of the $220 billion server CPU market by 2030, while accelerating Helios AI rack deployment, expanding its Venice server CPU line, and upgrading the ROCm software platform. The company added that supply chain capacity remains its main operating constraint and said it is working with TSMC and other partners to secure more advanced-node output.

Altimeter Capital partner Clark Tang argues that Nvidia is no longer just a chip supplier to the AI industry. In his view, the company has been building the two pillars that historically defined hyperscalers: an operating layer that abstracts and manages infrastructure, and a financing layer that funds capacity ahead of demand. Tang says this combination is turning Nvidia into a “synthetic hyperscaler,” one that is starting to displace Amazon, Microsoft, and Google in parts of the AI compute supply chain. His thesis begins with a shift in infrastructure economics. Traditional hyperscalers built strong margins by converting enterprise capex into opex and using software to maximize utilization of shared hardware. Tang says AI workloads break that model. Large-scale training depends on tightly synchronized GPU clusters, while inference is highly sensitive to tokens per watt and time to first token. In that setup, virtualization and networking layers that worked well in the cloud era can become a drag on GPU performance. He also points to the rise of neocloud providers, which offer lower-margin, AI-focused infrastructure but often lack the balance sheet strength to finance aggressive buildouts. Tang says Nvidia has moved to close that gap with software such as DSX OS, Mission Control, Omniverse, and Dynamo, while also standardizing hardware and bringing in third-party capital from firms including Apollo, BlackRock, Blackstone, Goldman Sachs, and KKR.

Intel’s shares came under pressure after the company expanded its planned common stock offering from $15 billion to $20 billion, pricing roughly 210.5 million new shares at $95 each. Reuters reported that the stock fell 4.1% on Aug. 10 after the original deal was announced, with dilution emerging as the clearest near-term concern for investors. Intel’s latest SEC filing said the company had about 5.044 billion shares outstanding as of July 17, making the base offering equal to roughly 4.2% of the existing share count, not including an underwriter option for up to 31.6 million additional shares. Intel said the proceeds will go to general corporate purposes, including capital expenditures and working capital, and estimated net proceeds of about $19.7 billion before any exercise of the additional share option. The fundraising comes as AI-related infrastructure demand lifts Intel’s server and manufacturing business. In the second quarter of 2026, Intel reported $16.1 billion in revenue, up 25% year over year. Data Center and AI revenue rose 59% to about $6.3 billion, while Intel Foundry revenue climbed 31% to $5.8 billion. The company has also raised its 2026 capital spending forecast from $18 billion to $20 billion as AI server demand runs above expectations.

Intel CEO Chen Liwu said during an appearance on the latest episode of the TechSurge podcast on Aug. 13 that demand for CPUs is enormous and that he fields daily calls from other chief executives who want more supply. To meet that demand, he said Intel is aggressively expanding CPU capacity and also developing new CPU architectures. Turning to memory, Chen said there are many ways to stack CPUs and memory, and that the industry has not innovated enough in memory, creating room for exploration. He acknowledged that he once avoided memory investments because he considered it a commodity business, but said that view has now changed. Chen also revealed that his friend Shock Lee, a former head of SK hynix, has joined, and he hinted that observers might guess what he is working on, while adding that nothing is ready to announce. He said he looks at short-, medium-, and long-term demand. On commitment, Chen said he typically stays at companies for 10 years — he cited Cadence — and that he is in it for 15 years at Intel. He told the board he focuses on a 10- to 15-year future and wants to build a platform that benefits the entire industry.