독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-22168 게시물

새로고침

ZK Research Chief Analyst Kerravala: Nvidia Is Shifting from Selling Chips to Selling an AI Factory Operating Model

주제 · 英伟达AI工厂战略중대
2026-10-22 04:52 GMT+8

Nvidia is shifting the competitive dimension of AI infrastructure from "whose chip is faster" to "who can get customers making money faster." ZK Research Chief Analyst Zeus Kerravala offered this assessment following the October 21, 2026, GTC conference in Berlin, Germany.

During his keynote address, Jensen Huang spent little time discussing silicon, instead redefining AI as a manufacturing sector: "We were early to observe that, in fact, AI is a manufacturing industry. It is no longer a tool used by us, but part of machinery, infrastructure, factories producing intelligence."

Behind this redefinition lies a specific economic calculation. On stage, Huang provided a rough estimate: a one-gigawatt AI factory costs about $50 billion and should generate approximately $40 billion in annual revenue. If deployment is delayed by one quarter, it results in a loss of about $10 billion in revenue. "$10 billion worth of revenue will make anybody anxious," he said. Kerravala described this approach as "interesting and correct," pointing out that the cost keeping AI builders awake at night is no longer hardware prices, but the cost of hardware sitting idle while crews pull cables. Modern AI construction is essentially an engineering project requiring kilometers of copper wire and fiber optics assembled on-site by scarce specialized labor, particularly in Europe.

Nvidia's response is the DSX POD: a complete Vera Rubin AI factory reference design pre-fabricated by partners in containers. Each POD contains seven Vera Rubin NVL72 racks, approximately 20 kilometers of NVLink copper cable, and 17 kilometers of expansion network cabling, all wired, plumbed, and tested before shipment. Nvidia states that deployment time drops from six months to as little as 14 days, assuming the site is ready.

However, Kerravala explicitly marked the boundaries: DSX POD will not replace gigawatt-scale campuses, which still benefit from economies of scale; the 14-day figure assumes the site is already prepared; and while containers compress integration, they do not solve grid interconnection queue issues. He wrote: "DSX POD moves the longest pole in the tent, but does not pull it out." Delivery is scheduled for early 2027.

In terms of the competitive landscape, Kerravala concluded that Nvidia's rivals are still primarily competing on chips, while Nvidia increasingly competes on "time to revenue," backed by an installed base where Volta GPUs from 2019 are still serving customers. For CIOs and operators weighing multi-billion-dollar AI investments, this durability combined with faster first-token time is harder to beat than any benchmark test.

It is important to note that Huang's figures of $50 billion/$40 billion are his own "back-of-the-envelope math" estimates, not audited financial data; Nvidia's claims regarding deployment times are vendor-reported and have not yet been independently verified. Kerravala's judgment as an analyst also constitutes personal interpretation.

출처:SiliconANGLE

리서치

Inception Labs Releases Mercury 3: Self-Tests Claim 4x to 8x Faster Than GPT-6 Luna

중대
2026-10-21 22:45 GMT+8

US AI startup Inception Labs released Mercury 3 on October 21, 2026, a language model based on diffusion technology. Unlike traditional autoregressive models that generate text word by word, Mercury 3 first creates a rough draft and then refines it step by step, allowing for parallel token processing.

According to The Decoder, Inception Labs claims in internal benchmarks that Mercury 3 generates over 1,000 tokens per second on Nvidia GPUs, running four to eight times faster than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5. It leads rivals in eight of 12 tests, particularly in instruction following (scoring 76.84 on IFBench) and autonomous coding (scoring 51.08 on SWE-Bench Atlas). However, Claude Haiku 5.5 wins the longer coding test DeepSWE, and GPT-6 Luna edges ahead in some Artificial Analysis tests.

The model supports a context window of up to 1 million tokens and offers an OpenAI-compatible API. Inception Labs states that Mercury 3 is designed for latency-critical applications such as autonomous AI agents, smart glasses search, and real-time customer support. Early partner Octen AI reported that median response time for smart glasses search dropped to 0.9 seconds with Mercury 3, noting that the language model was previously the bottleneck.

It is important to note that no independent third-party verification of these performance claims exists yet. Inception Labs indicates that Mercury 3 has more than 40 times the parameters of its predecessor Mercury 1 but runs at similar speeds, suggesting that diffusion models scale effectively like standard language models. The exact parameter count remains undisclosed.

출처:The Decoder

리서치

US Startup Recursive Releases Flint: A Laptop Coding Model It Says Was Largely Built by AI

중대
2026-10-22 04:27 GMT+8

US-based AI startup Recursive released its first public model, Flint, on October 21. The company describes Flint as an open-source coding model designed for laptops, stating that its development was "largely built with the help of AI."

A key feature of Flint is the tool used to create it: Recursive's "Eureka Machine." According to TechCrunch, this automated research system contributed to nearly every step of the model's creation, from pre-training to post-training, enabling a small team to build the model from scratch in under four months. This release serves as an early public test of "recursive self-improvement," the concept of using AI to build better AI.

Technically, Flint is a mixture-of-experts model with 40 billion total parameters, but only 2.5 billion are active for any given task. This architecture makes it lightweight enough to run on ordinary laptops. In contrast, Meta's Muse Spark Open, an open model released last week for Macs, uses 816 billion parameters and requires at least a terabyte of memory.

Recursive's benchmark data indicates that Flint generates approximately 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, roughly twice the speed of similar models like Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning on the same hardware. On SWE-Bench Verified, a benchmark measuring real-world software bug fixing, Flint outperformed other models under 50 billion parameters. However, these performance figures are vendor-reported and have not yet been independently verified at scale.

Flint is available for free download on Hugging Face under the Apache 2.0 license, allowing for commercial use and modification. Because the model runs locally on user devices rather than cloud servers, developers can work offline, maintain code privacy, and avoid usage fees associated with AI cloud services.

While Recursive acknowledges that Flint requires more guidance than large paid coding assistants for complex tasks, early reviews have been positive. Wired reporter Will Knight used Flint on a Mac with 48GB of memory to build several small programs, noting that it handled coding requests accurately and faster than other smaller models.

출처:TechCrunch

리서치

Blogger Peter Yang tests OpenAI's Dot: How an always-on agent acts as a 'Chief of Staff'

주제 · OpenAI智能体Dot중대
2026-09-30 20:20 GMT+8

OpenAI released its always-on personal agent, Dot, on September 30, featuring an independent cloud computer running 24/7. Since the product is not yet fully launched, overseas blogger Peter Yang showcased 10 real-world cases in a beta test video, revealing Dot's capabilities and design logic in actual work scenarios.

Peter Yang believes Dot is not just a chatbot but a 'Chief of Staff' designed for work contexts. In the most impressive case involving sponsorship revenue audits, Dot proactively reviewed company accounts and found five issues, including unbilled Riverside contracts and a $10,000 discrepancy in Google accounts. This ability to actively dig through scattered information and identify anomalies is seen as Dot's biggest value compared to traditional on-demand tools.

In another fitness app development task, Dot did not write code or generate images itself. Instead, it broke down the task: delegating code-related needs to Build threads, calling image capabilities to generate screenshots, and coordinating bug fixes. From this, Peter Yang summarized the ideal usage pattern: let Dot be the manager and the user be the director, with Dot responsible for tracking progress, reporting results, and dispatching other ChatGPT threads.

However, the tests also exposed significant shortcomings. In one newsletter writing task, Dot failed due to Substack login authentication issues but did not clearly inform the user of the failure, making it difficult to locate the bottleneck when multiple threads were running. Additionally, while Dot aims to be a unified command entry point amidst increasingly complex ChatGPT features, it may also exacerbate cognitive load for ordinary users. Peter Yang noted that if even daily AI tool users find the interface complex, ordinary users may struggle even more to find the functions they need.

출처:zhidx.com

리서치

Meta Reports Q3 2026 Results: Revenue Up 26%, Plans Non-Voting Class C Stock for AI Funding

주제 · Meta千亿AI投入策略구조적
2026-10-22 04:01 GMT+8

U.S. social media giant Meta Platforms (Nasdaq: META) released its financial results for the third quarter ended September 30, 2026, on October 21. The report shows revenue reached $64.60 billion, up 26% year-over-year, while net income surged 264% to $9.86 billion. This profit growth was primarily driven by a drop in the effective tax rate from 87% to 24%, which partially offset a $9.89 billion charge related to a settlement with U.S. state attorneys general over teen mental health litigation. Excluding this one-time expense, adjusted operating income would have been $22.88 billion, and adjusted diluted EPS would have been $7.40.

In terms of capital allocation and governance, Meta's Board of Directors approved the establishment of a new class of publicly listed, non-voting Class C common stock, effective January 2027. Recommended by a Special Committee of independent directors, this move aims to provide financing flexibility for massive AI infrastructure investments while ensuring that Founder and CEO Mark Zuckerberg's voting control remains undiluted. Meta expects full-year 2026 capital expenditures to range between $135 billion and $145 billion, reflecting continued aggressive investment in 'personal superintelligence' initiatives.

Meta's CFO noted that AI investments are already delivering substantial incremental advertising revenue through stronger engagement and better monetization efficiency. While losses at the Reality Labs segment are expected to narrow to $18-19 billion compared to 2025 levels, the company is reallocating resources toward priority areas, primarily AI. With the introduction of the new share class, investors should monitor how Meta balances long-term strategic autonomy with the cash flow pressures of sustained high-capex spending.

출처:prnewswire.com

리서치

China's Ministry of Commerce Proposes to Include AI Models in Export Controls: The Regulatory Gate Behind Open-Source Automatic Licensing

주제 · 中国AI出口管制草案구조적
2026-10-21 21:43 GMT+8

Chinese open-source AI models have long been viewed as Beijing’s response strategy to Washington’s chip blockade: exchanging public weights for global adoption rates and brand influence. On October 16, 2026, the Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology (MOST) of the People's Republic of China jointly released Draft Announcement No. 48 of 2026, proposing for the first time to include AI models themselves in the Catalogue of Technologies Prohibited or Restricted from Export. After spending three days reading through the draft line by line, ChinaTalk author Quinn Ennis pointed out that the core intent of this document is not to stop open source, but to insert an institutional checkpoint before open-source releases.

The draft sets three thresholds for frontier models: training floating-point operations reaching 10^26 FLOPs, total parameters reaching 2 trillion, or capabilities reaching a globally leading level.

Among these three thresholds, the parameter count is the item with the most practical constraint on current Chinese models. According to data from Epoch AI, since mid-2024, only four Chinese models—Kimi K3, ERNIE 5.0, Qwen 3.8 Max, and Qwen3.8-2.4T-A95B—have exceeded 2 trillion parameters. Models from DeepSeek, Meituan, Xiaomi, and Z.ai are near the trillion-parameter mark, and the author believes their next versions are likely to cross this line. As for the 10^26 FLOPs threshold, no released Chinese model has reached it as of October 20, 2026; the largest training compute power is approximately 2×10^25 FLOPs.

What truly deserves attention is not the thresholds themselves, but the actual operation of the open-source exemption. The draft's Q&A states that open-source AI technologies apply the "immediate submission, immediate license issuance" automatic licensing principle, but the Q&A leaves room for actual delays with the phrasing "basically achieve" (基本实现).

During the transition period before formal rules take effect, the draft requires pre-export registration for all exports involving technologies added to the new catalogue, including exports that have already begun but are not yet completed.

The author specifically cited the Manus case as institutional background. In December 2025, Meta signed an agreement to acquire the Chinese agent startup Manus, which had previously relocated to Singapore. Following MOFCOM review, a security review led by the National Development and Reform Commission (NDRC) resulted in the deal being reversed. Ennis noted that the third registration trigger condition in the draft—changing the ownership or actual controller of domestic technology to a foreign entity—essentially codifies the jurisdictional logic of the Manus case into the rules, preventing companies from evading Chinese judicial jurisdiction through re-domiciliation.

Regarding agent export controls, the draft's triggering capabilities include autonomous planning and task decomposition, multi-agent collaboration, and tool use. However, the author points out that these capabilities are not unique to Chinese enterprises, and the various thresholds (such as integrating 200 third-party tools) appear somewhat arbitrary. What is truly controlled is the specific architecture supporting Chinese agent products in reaching these thresholds at lower costs.

출처:chinatalk.media

리서치

Zain Asgar, Founder of AI Inference Orchestration Company Gimlet Labs: Why Specialized Accelerators Are Making a Comeback as General-Purpose Chips Stop Getting Faster Automatically

중대
2026-10-22 03:51 GMT+8

In 1990, Eugene Brooks of Lawrence Livermore National Laboratory asserted at a supercomputing conference that nothing could withstand the attack of 'killer micros'—where 'micro' referred to microprocessors, meaning general-purpose chips that performed well on most types of computation. They were also being sold by the millions per year into desktop PCs, delivering faster and faster performance each year.

The victory of general-purpose chips was not because specialized chips did anything wrong. Asgar points out that general-purpose microprocessors actually outperformed them even on the target workloads for which the specialized chips were designed. The ideas behind vector processors were later absorbed into CPUs as SSE and AVX instructions, and massively parallel designs influenced multi-core CPU and GPU architectures—the legacy of specialized companies lives on inside general-purpose chips, but the companies themselves disappeared.

This virtuous cycle was supported by three mutually reinforcing elements: generality drives volume—microprocessors entered desktop PCs with shipments of tens of millions, while specialized hardware had much smaller volumes because it ran only one type of workload; volume drives performance—huge shipments brought lower costs and more R&D funding, with single-processor performance growing over 50% annually between 1986 and 2003; and performance in turn reinforced generality—faster clock speeds unlocked applications that were previously unfeasible, such as dedicated video decoding cards being eliminated within two years after the appearance of the Pentium II.

The first cornerstone of Dennard scaling—the proportional reduction of voltage in smaller transistors—failed around 2005. Frequencies therefore stagnated, and multi-core became a compensatory measure, but not all applications can be efficiently parallelized. Around 2012, the cost per transistor also hit bottom: the growth in fab and process costs exceeded the growth in the number of transistors per wafer. The memory wall was named as early as 1995 and remains unsolved—David Patterson estimates that NVIDIA GPU compute power grew 80-fold from 2012 to 2022, while memory bandwidth grew only 17-fold, because compute scales with chip area, whereas bandwidth is limited by chip perimeter.

Asgar felt this shift earlier during his PhD studies in Mark Horowitz’s group at Stanford. His co-founder Omid Azizi, working with Horowitz, found that a 720p video encoder running on a general-purpose chip was about 500 times less energy-efficient than running it on a specialized ASIC. When the free lunch disappeared, the energy efficiency advantage of specialized chips became meaningful again.

GPUs themselves are survivors of specialized chips. The GeForce 256 in 1999 did one thing: rendering triangles for game graphics. But unlike vector processors and Lisp machines, it had a mainstream audience and shipments of millions. After frequency stopped scaling in 2005, chips had to shift from doing the same thing faster to doing more things simultaneously. The design of GPUs, with thousands of simple cores and extremely high memory bandwidth, happened to align with this shift, making them an ideal platform for large-scale parallel workloads ranging from early machine learning to cryptocurrency to high-performance computing.

출처:a16z.news

리서치

Insilico Medicine and Servier Achieve First Milestone in AI-Driven R&D Collaboration, Triggering Undisclosed Payment

주제 · 英矽智能与施维雅合作중대
2026-10-21 23:41 GMT+8

Insilico Medicine has published a video on its official channel confirming substantive progress in its multi-year, AI-driven research and development collaboration with French pharmaceutical company Servier. According to the video description, two oncology programs under this partnership have initiated and completed target nomination, marking the first R&D milestone achieved by Insilico using its proprietary Pharma.AI platform.

The attainment of this technical node directly triggered an undisclosed milestone payment from Servier to Insilico. The collaboration was announced in early 2026, covering multi-target oncology areas, with a total potential value of up to $888 million, including $32 million in upfront and near-term R&D payments. For readers tracking the commercialization of AI in drug discovery, this event provides direct evidence of a major pharmaceutical company making staged payments based on specific AI technical outputs.

출처:youtube.com

리서치

Airbnb CTO Ahmad Al-Dahle on Building a 'Company Brain' for AI Agents

중대
2026-10-22 03:09 GMT+8

For teams exploring the integration of AI agents into internal enterprise processes, the biggest hurdle is often not model capability itself, but enabling AI to understand the tacit institutional knowledge that experienced employees rely on but never documented. The latest episode (781) of the TWIML AI Podcast addresses this pain point by featuring Ahmad Al-Dahle, Chief Technology Officer of US-based home-sharing platform Airbnb, who delves into how his team is solving this problem by building a 'company brain'.

In the episode, Al-Dahle introduces Project Everest, an internal system designed to automatically capture knowledge from employees' workflows for use by AI agents. The content goes beyond technical implementation to touch upon deeper challenges in knowledge management: How to prevent 'context poisoning' and interference from stale information as people and agents continuously add new data? Why can too much context actually backfire? And how to establish effective evaluation mechanisms to determine whether shared knowledge truly improves engineering efficiency?

Additionally, the discussion covers methods for measuring AI productivity, advocating for a shift beyond simple metrics like token consumption or pull request counts towards more substantive output assessments. Al-Dahle also shares Airbnb's practical thoughts on the limitations of chat-based interfaces and the challenges of deploying support agents in a marketplace environment where resolving disputes requires mediating between conflicting interests. For engineering leads and product managers focused on enterprise-level AI adoption, particularly those interested in knowledge graphs and agent collaboration, this serves as a valuable record of industry practice with specific reference points.

출처:twimlai.com

리서치

For Context Ep. 1: Orbit's Max Tatton-Brown on Capturing Meeting Context with Granola

간단한 요점
2026-10-21 23:03 GMT+8

For Context is a podcast exploring how people use the Granola MCP to give their agents context and get work done. In its first episode, it features Max Tatton-Brown from Orbit, focusing on capturing valuable ideas from meetings.

According to the official episode description, Max Tatton-Brown discusses the importance of retaining meeting insights, stating: "If you can capture them, you can piece together something really valuable." This highlights the potential of technical solutions in preserving spontaneous thoughts during discussions.

The episode also covers specific scenarios for connecting Granola MCP to various workflows, such as creating agendas with Claude or handling follow-ups in Slack. For professional readers interested in how AI agents integrate unstructured meeting data to improve efficiency, this resource offers a clear perspective on product application.

출처:youtube.com

리서치

Walmart to Open Site to AI Agents Next Month via MCP, Diverging from Amazon

주제 · 沃尔玛AI代理购物중대
2026-10-22 03:30 GMT+8

US retail giant Walmart announced that it will allow users to connect their own AI agents to its website starting next month. This move enables automated shopping within strict safety boundaries while keeping final control in the hands of consumers.

According to reports, Walmart has built an "Agent Gateway" based on the Model Context Protocol (MCP), an open standard that allows AI agents to plug into web services structurally. This approach avoids the technical hiccups agents often face when navigating traditional browsers. The gateway ensures agents only access information necessary to complete tasks, preventing them from viewing broader shopping histories or personal data. For instance, if a customer frequently buys sourdough bread, the system helps the agent find those items without sharing the customer's full purchase history.

To ensure security, each AI agent gets its own profile within the user's Walmart account. By default, all purchases made by agents require explicit user approval, though customers can set budgets (e.g., up to $100) for autonomous execution. Users can also review agent transactions separately and disable the agent at any time. Walmart stated that agents must not change payment methods or take actions beyond the shopper's intent, and non-compliant agents will be blocked.

This strategy stands in stark contrast to Amazon, which recently blocked shopping agents from Perplexity and Meta's Muse AI assistant. By embracing AI agents under controlled terms, Walmart aims to capture the emerging "agentic commerce" market. The feature will also extend to its Sam's Club warehouse stores.

출처:The Verge

리서치

Sources: Meta Plans Non-Voting Shares to Fund AI Buildout Without Diluting Zuckerberg

주제 · Meta无投票权新股融资구조적
2026-10-22 03:17 GMT+8

According to an exclusive report by Semafor, Meta Platforms (the US social media giant) plans to introduce new non-voting stock, allowing it to fund its artificial intelligence infrastructure buildout through share sales without diluting the voting control of CEO Mark Zuckerberg.

The plan could be announced as soon as Wednesday when Meta reports quarterly earnings. The new equity structure would enable the company to issue stock for acquisitions, to attract top AI talent, and to pay for chips, while maintaining the founder's super-voting position. This strategy mirrors Alphabet's (Google's parent company) move in 2014, which introduced non-voting shares alongside super-voting shares to keep co-founders Larry Page and Sergey Brin's combined control above 50%, facilitating the issuance of $85 billion in new stock earlier this year for AI investments.

Big tech companies have historically balanced new stock issuance by buying back shares in the market. However, they have paused repurchases as the cost of their AI investments mounts, now finding themselves as net issuers of shares, which risks diluting insiders' stakes. A Meta spokesperson declined to comment.

출처:semafor.com

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

The author also identified a wording issue worth noting in the draft: the restriction condition for multi-agent collaboration technology is "overall task completion time ≤ 2 times the minimum completion time of a single agent." The author questions whether this wording is intentional, as it allows a multi-agent system to be twice as slow as a single agent.

In terms of jurisdictional allocation, the draft stipulates that export licenses are reviewed by the provincial commerce authority where the exporter is located, in conjunction with the provincial science and technology authority, rather than being directly approved by MOFCOM. However, Ennis cited the 2020 ByteDance TikTok sale application and the 2023 CATL export of battery technology to Ford cases to illustrate that central authorities effectively retain final decision-making power in cases involving the highest national security interests.

The timing of this draft's release also drew the author's attention. The public comment period ends on November 15, just three days before the alleged November 18 meeting between Chinese President Xi Jinping and U.S. President Donald Trump at APEC. Ennis believes that while the final text will not be finalized within these three days, the options to tighten, narrow, or shelve the draft itself can serve as bargaining chips.

For global AI developers and adopters, Ennis’s analysis points to a specific judgment: the global appeal of Chinese open-source models may decline due to licensing uncertainty—if licenses can be revoked at any time, adopters might turn to non-Chinese models.

Finally, it must be clarified that this is a legal stage issue. Announcement No. 48 is a draft, not an effective regulation. The public comment period runs from October 16 to November 15, 2026. The author himself reminded readers that during the catalogue revisions of 2022–2023, public comments and expert consultations led MOFCOM to delete and narrow several proposed entries. Before the finalization of the 2024 Dual-Use Items Export Control Regulations, MOFCOM reviewed nearly 600 opinions. The specific threshold definitions, open-source exemption boundaries, and licensing execution mechanisms in the final text may change substantially.

Bill Dally of NVIDIA broke down the sources of roughly 1,000x improvement in single-GPU inference performance over ten years at Hot Chips 2023: about 16x came from simpler number formats (such as FP32 to BF16), 12.5x from new instructions, 2x from leveraging sparsity, and process improvements from 28nm to 5nm contributed only 2.5x. This means that process advancements, which drove industry performance growth for forty years, now contribute less than other factors. Asgar thus proposes an interesting reversal: perhaps we should now call GPUs general-purpose accelerators and CPUs specialized chips.

Back to the present, inference demand is exploding. Asgar cites data from a 2026 Goldman Sachs report: token demand will grow 24-fold from 2026 to 2030, while available power will grow only about 2-fold. Under conditions where physical infrastructure remains scarce, efficiency per watt becomes the most critical competitive dimension. Inference itself is a highly heterogeneous workload—the prefill and decode phases benefit from different hardware trade-offs. Selecting the correct chip architecture for each segment of the workload can bring significant performance improvements, even if underlying resources no longer scale automatically as they did in the past.

Asgar believes there are three key differences between today and the 1990s: it is much easier to write optimized code for new accelerators using AI agents; chip design infrastructure is more mature (most chip companies do not need to build their own fabs); and the scale and power constraints of inference workloads far exceed those of the past. His company, Gimlet Labs, splits inference workloads across different chips such as NVIDIA GPUs, LPUs, Google TPUs, Cerebras wafers, d-Matrix, Trainium, and AMD GPUs, claiming that heterogeneous splitting technology is 2 to 10 times faster than homogeneous stacks under the same power envelope.

This conclusion comes with a boundary acknowledged by the author himself: GPUs are following the path of CPUs in the past, directly internalizing the concept of inference-specific hardware—adding Tensor Cores in 2017, and MXFP4 and NVFP4 in 2024. If GPUs continue to absorb specialized features, the window for external specialized chips may close again. Asgar expects that the pendulum will eventually swing back to general-purpose architectures, but the era of specialized chips has arrived; even if general-purpose chips win again in the future, today's specialized hardware may become tomorrow's mainstream, with its contributions continuing to live on in general-purpose chips.

A note on conflicts of interest: Asgar is the Co-founder and CEO of Gimlet Labs, whose main business is heterogeneous inference orchestration. The data regarding the 2 to 10x performance improvement from splitting technology comes from Gimlet's own analysis (labeled 10/20/26) and is not independently verified by third parties. The article also includes recruitment and promotion information for expanding its inference cloud. These do not negate the logical value of the historical analogy and physical constraint arguments, but readers evaluating whether heterogeneous orchestration can truly deliver commercial advantages should view Asgar's optimistic judgment as a claim from a stakeholder.