LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-10-28154 publicaciones

Actualizar

SemiAnalysis Podcast Discusses Anthropic S-1: Inference Revenue Per Megawatt Nearly Quadrupled

Tema · AnthropicMaterial
2026-10-28 04:41 GMT+8

SemiAnalysis released a podcast episode on its YouTube channel focusing on key financial metrics from AI company Anthropic's latest S-1 filing. The episode description cites specific data: Anthropic's "inference revenue per megawatt" across its entire fleet increased from $16 million last year to over $60 million currently. This metric is used to support the claim that "AI model businesses scale extremely well."

The episode also notes that for the three months ended June 30, Anthropic's revenue grew 1,368% year-over-year, while compute and infrastructure expenses grew only 300% during the same period. The hosts further explore whether this efficiency gain stems from Anthropic's proprietary advantages or broader market hardware improvements, pointing out that GB series chips are significantly better than Hopper architecture. This episode offers readers interested in AI unit economics and infrastructure ROI a concrete entry point into these discussions.

Fuentes:youtube.com

Investigación

NVIDIA Releases GPU Fractioning Operator: Integrating CUDA 13.4 VRAM Isolation Capabilities into Kubernetes

Material
2026-10-28 04:50 GMT+8

Many AI inference and fine-tuning tasks require only a small fraction of a GPU's memory, yet allocating an entire card to each workload leaves significant capacity idle. In a technical blog post dated October 27, 2026, NVIDIA introduced a new underlying capability added in CUDA 13.4: memory boundary enforcement that is transparent to applications. When a workload declares a specific memory size, CUDA restricts it to that range; allocations exceeding the limit fail, as if running on a smaller GPU. This offers a guaranteed minimum value plus an expandable upper limit when memory is available. While CUDA provides this capability, Kubernetes clusters cannot invoke it directly—nodes must run MPS daemons, write limits and compute modes before container startup, and report usage by fraction rather than by whole GPU.

The GPU Fractioning Operator, paired with the NvFractions mode of KAI Scheduler v0.18.0, bridges this gap: the scheduler selects GPUs and writes per-container annotations, while the operator executes configurations on the node. There are two compute modes: time-slicing allows workloads to take turns using the entire GPU, each maintaining its own independent error domain; sm-sharing allocates streaming multiprocessors (SMs) simultaneously to multiple workloads, sharing a single error domain. A single GPU runs only one compute mode at any given time, and KAI Scheduler tracks this constraint per GPU rather than per node.

Fuentes:NVIDIA Technical Blog

Investigación

GM Plans Over $2 Billion Investment in Lansing Plant for Hybrid Production, Restructures Federal Grants

Tema · 通用汽车兰辛工厂改造Material
2026-10-28 05:47 GMT+8

General Motors stated on Tuesday that it plans to invest more than $2 billion in the broader Lansing region of Michigan over the next several years, including $1.25 billion to prepare its Lansing Delta Township Assembly plant for future hybrid vehicle production. This move marks a significant adjustment in capital allocation strategy by the Detroit-based automaker as it responds to shifting consumer preferences toward hybrids amid declining sales volumes.

The company said it is working with the U.S. Department of Energy to restructure a Domestic Manufacturing Conversion Grant previously awarded under the Biden administration. The original $500 million grant was intended to support converting a different Lansing factory to produce electric vehicles, including battery-electric, plug-in hybrid, and hydrogen fuel cell models. According to GM, those funds were never released and will now be redirected to hybrid production at the Delta Township facility.

Currently, the Lansing Delta Township plant assembles gas-powered Buick Enclave, Chevrolet Traverse, and GMC Acadia models, producing a record 291,000 vehicles in 2025. While GM has not provided specifics on when hybrid vehicles will begin production and sales, insiders told TechCrunch to expect hybrids across multiple brands rather than a single vehicle line. A Reuters report, citing anonymous sources, indicated plans to build hybrid versions of the Chevrolet Traverse, Buick Enclave, and GMC Acadia at the factory.

Hybrid vehicles have become increasingly popular in the U.S. due to rising gas prices. Toyota leads the market in hybrid sales, followed by Honda. Ford also gained considerable market share, selling approximately 228,000 hybrids in 2025. In contrast, GM had previously bet heavily on battery-electric vehicles, investing billions to shift its portfolio away from internal combustion engines. However, this strategy changed under the Trump administration with the end of federal EV tax credits and slower-than-expected demand. GM continues to sell numerous battery-electric vehicles but is now pivoting its product mix toward hybrids.

Fuentes:TechCrunch

Investigación

Robinhood Q3 Profit Falls 14% as Crypto Revenue Slumps Despite Prediction Market Growth

Tema · RobinhoodMaterial
2026-10-28 04:40 GMT+8

US online broker Robinhood Markets (HOOD) reported its third-quarter financial results on October 27, 2026. Although record trading volumes in stocks, options, and prediction markets drove total revenue up 12% year-over-year to $1.42 billion, net income fell 14% to $476 million. Diluted earnings per share came in at $0.51, missing analyst estimates of $0.57. Shares dropped approximately 4% in after-hours trading following the release.

Shifts in revenue composition were central to the report. Crypto trading revenue, previously a key growth driver, plunged 55% year-over-year to $120 million due to slower market activity. Net interest revenue also declined 9% to $415 million, attributed partly to reduced securities lending. Conversely, emerging segments performed strongly: event contract (prediction market) revenue surged to $173 million, up 11% quarter-over-quarter and more than six times higher than a year ago, while options trading revenue rose 31% to $399 million.

Rising costs and non-operating losses further compressed margins. Operating expenses increased 16% to $739 million, and provisions for credit losses nearly tripled to $71 million as the company expanded its credit card business. Additionally, a $67 million unrealized investment loss from the Robinhood Ventures Fund I reduced earnings per share by approximately 7 cents. CFO Shiv Verma stated that the core business remains robust, noting strong momentum entering Q4 with over $13 billion in net deposits already recorded in October, including contributions from the Dell family to Trump Accounts.

Fuentes:Yahoo Finance

Investigación

Ben Carlson: The Smarter You Are, the More Likely You Are to Make Big Mistakes in Investing—How Self-Deception Amplifies Blind Spot Bias

Material
2026-10-28 03:11 GMT+8

In 2015, then-US Vice President Joe Biden visited the Silicon Valley blood-testing company Theranos and stated during a roundtable discussion that he had just witnessed "the laboratory of the future." The company promised to perform a full suite of diagnostics with just a drop of blood from a patient's fingertip. Founder Elizabeth Holmes claimed this was "the most important thing humanity has ever built." The entire company was a fraud—the lab shown to visitors was a staged set, the equipment did not actually work, and most employees were told to stay home that day. Biden was far from the only big name to be conned by Holmes.

The Walton family, heirs to the Walmart fortune, lost $150 million. Media mogul Rupert Murdoch invested $125 million based strictly on gut feeling alone; he later sold the entire stake for $1 to take a tax write-off for the enormous loss on a worthless investment. Betsy DeVos, Education Secretary under President Trump, and her wealthy family lost $100 million. Partner Fund Management, a well-known hedge fund, purchased shares at a valuation that pegged the company’s value at $9 billion, making Holmes worth a staggering $5 billion on paper.

US wealth manager Ben Carlson revisited a chapter from his book Don’t Fall For It in his blog A Wealth of Common Sense, attempting to answer an counter-intuitive question: Why were these smart, wealthy, and top-educated individuals deceived? His answer was not that they weren't smart enough, but that they were too smart. The core mechanism is blind spot bias—the tendency for people to see bias in others more easily than in themselves. Carlson argues that smart people suffer from this more severely than anyone else because they possess stronger self-deception capabilities. Once a person has achieved sufficient success, they tend to attribute it entirely to their own abilities rather than opportunity, timing, or luck. This psychological tendency to "personalize success" leads them to believe they won't make mistakes in unfamiliar fields, while those around them, respecting their intellectual status, do not question their judgment.

This argument is supported by empirical evidence. Researchers Richard West, Russell Meserve, and Keith Stanovich tested seven types of cognitive abilities and found that high cognitive ability does not attenuate metacognitive bias—in other words, smarter people are worse off at recognizing their own thinking blind spots. Crucially, those who were aware of their own biases were not better able to overcome them. Carlson concludes that smart people may be less capable than ordinary people at identifying their own limitations because they have more "firepower" to convince themselves through self-deception.

For readers, the practical implication of this mechanism is: when making investment decisions in unfamiliar fields, analytical ability is not a safety cushion but a risk amplifier. When you let your guard down because "I understand complex systems," that is precisely when blind spot bias is most active. The Theranos board included former US Secretary of Defense Jim Mattis, former Secretary of State George Shultz, retired US Navy Admiral Gary Roughead, former Secretary of State Henry Kissinger, and former Secretary of Defense William Perry. These individuals' judgment was unquestionable, yet it was this collective "smart people endorsing each other" that allowed the fraud to go unchallenged for years.

Fuentes:A Wealth of Common Sense

Investigación

Meta Publishes 2026 Python Typing Survey: AI Usage for Type Checking Rises to 38.4%, Nearly Half of Developers Mix Multiple Checkers

Material
2026-10-27 23:59 GMT+8

Meta's engineering team published the "2026 Python Typing Survey" report on October 27, 2026, based on 1,794 responses, revealing the latest practices regarding the integration of static type checking and AI tools among Python developers. The data shows that the percentage of respondents who believe "AI makes generated code more accurate/reliable" jumped from 18.5% in 2025 to 38.4% in 2026. In specific workflows, 42.5% of respondents use AI tools to write type hints, while 32.6% manually write types to improve AI-generated code. Overall, 61.0% of respondents reported using AI tools in at least one typing-related activity.

Despite increased AI integration, developers face significant toolchain fragmentation. The survey indicates that nearly half (49.9%) use two or more type checkers simultaneously, up from 46.0% last year. Speed was the primary reason for choosing a checker (56.5%). Notably, some respondents pointed out "lazy" behaviors in AI coding agents, such as overusing object types or adding #type: ignore comments to suppress checker errors, raising new concerns about the quality of AI-assisted code.

Missing type information in third-party libraries remains the most common obstacle, cited by 65.1% of respondents. Dynamic features in popular libraries like Django, SQLAlchemy, Pydantic, NumPy, and pandas make it difficult for current type systems to fully describe their runtime behavior. Meta's Pyrefly team is working with Quansight to improve typing support for these libraries and is studying how AI coding tools interact with type checkers in agentic workflows.

Fuentes:Engineering at Meta

Investigación

SemiAnalysis Podcast Ep. 041: Decoding Anthropic's S-1 Filing and Compute Economics

Tema · Anthropic上市Material
2026-10-27 23:00 GMT+8

SemiAnalysis published Episode 041 of its podcast, titled 'EMERGENCY: Inside the Anthropic’s S-1 (Tokenomics/ Markets)', on October 27, 2026. Hosted by Jordan Nanos (@JordanNanos), Doug O'Laughlin (@fabknowledge), Jeremie Eliahou Ontiveros (@JeremieEO), and Joey Brookhart (@SaasquatchC), the episode aims to break down the economic logic and market implications behind Anthropic's recent S-1 filing with the U.S. Securities and Exchange Commission.

According to the episode description, core discussion points include Anthropic's 5.2x increase in revenue per dollar of compute since Q1 2025, and the trend where customers who spent $1 last year now spend $9. Additionally, the guests will analyze incremental gross margins running near 90%, half-trillion-dollar compute commitments, risks associated with customer concentration, the 2027 compute roadmap, and whether competitor OpenAI can keep up.

For professional readers interested in AI industry investment and the financial health of foundation model companies, this episode offers an immediate interpretive perspective based on public disclosure documents. While specific financial data and analytical conclusions require actual listening, the listed topics cover critical dimensions from profitability and datacenter views to new funding dynamics, helping to understand the operational efficiency and strategic considerations of leading AI companies during this capital-intensive phase.

Fuentes:youtube.com

Investigación

Anthropic Files for Nasdaq IPO: $80.7 Billion Loss in First Half of 2026, Lease Liabilities Surge Sixfold to $321 Billion

Tema · Anthropic上市Material
2026-10-28 04:49 GMT+8

Anthropic filed an initial public offering (IPO) application with the U.S. Securities and Exchange Commission (SEC) in June and officially announced its plan to go public on October 26, intending to list on the Nasdaq under the ticker symbol "ANT".

The trajectory of losses is the first set of numbers to attract attention: The company reported a full-year loss of $8.3 billion in 2024, a full-year loss of $41.9 billion in 2025, and a loss of $80.7 billion already in the first half of 2026. As of June 30, Anthropic held $112.4 billion in cash and cash equivalents. With 5,300 employees, this translates to revenue per employee of $6.6 million. International revenue accounted for 27.8% of total revenue, and no single customer contributed more than 10% of revenue—Elfenbein contrasts this with Nvidia, which has five clients each accounting for more than 10% of its revenue.

The most striking structural finding lies in the balance sheet. Anthropic is described as a typical 'asset-light' company, owning just over $1 billion in property and equipment (of which $432 million is related to finance leases). However, its future undiscounted lease obligations jumped from $50.2 billion at the end of 2025 to $321 billion. Elfenbein argues this reveals the true cost commitments behind the 'asset-light' label: the company does not own data centers but has locked into unprecedented long-term lease agreements.

Regarding corporate governance, Founder LLC holds 50.1% of the shares. This LLC has only seven members: founders Dario Amodei, Daniela Amodei, and five executives. The prospectus names only two competitors—OpenAI and Google—and lists 90 pages of risk factors, including legislative restrictions on data center construction in various parts of the United States and supply chain risks due to the high concentration of chip manufacturing at Taiwan Semiconductor Manufacturing Company (TSMC).

Elfenbein's core judgment is that Anthropic will use its high post-IPO stock price as 'cheap currency' to acquire competitors. He wrote, 'The benefit of an elevated share price is that it can be used as cheap currency to buy out your rivals.' He expects the company to use the raised funds to aggressively acquire competitors and explicitly stated that he would not participate in this IPO investment but would watch closely. This acquisition thesis is Elfenbein's personal prediction; the prospectus only states that funds will be used for 'general corporate purposes' and 'may be used to acquire other companies'.

Evidence boundaries to retain: All financial figures above are derived from Elfenbein's retelling of the S-1 document; this draft has not independently verified the original documents. The 'consensus valuation of $1.5 to $2 trillion' is Elfenbein's generalized description of market expectations, not a confirmed offering price. Details such as SpaceX charging Anthropic approximately $3.1 billion per month for GPU computing power, and Amazon having invested $13 billion since 2023 with potential additional investments of $20 billion, similarly rely on Elfenbein's unilateral retelling.

Fuentes:cws.substack.com

Investigación

Anthropic's IPO Filing Highlights Multi-Chip Compute Strategy, Analysts Warn of Threat to Nvidia Pricing Power

Tema · AnthropicMaterial
2026-10-28 03:31 GMT+8

US-based AI startup Anthropic has explicitly cited its ability to manage "a single, fungible pool" of compute resources from multiple strategic partners as a core competitive advantage in its initial public offering (IPO) application. According to Business Insider, the S-1 filing released Monday reveals that Anthropic currently utilizes hardware from four distinct chip families: Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs.

The company states that its software directs workloads to different chips based on productivity levels, noting that building this flexibility "took years and significant investment." This disclosure signals to investors a strategic intent to reduce dependence on any single supplier.

Patrick Moorhead, founder and CEO of Moor Insights & Strategy, pointed out that the real exposure lies in Nvidia's pricing power, stating, "A customer that can shift work across four platforms negotiates very differently than one that can't." Ashish Nadkarni, an analyst at IDC, added that the multichip approach also insulates Anthropic from fluctuations in Nvidia's chip supply.

However, breaking Nvidia's grip remains challenging. Moorhead emphasized that switching between chips is still difficult, particularly for training AI models rather than running inference, and only the largest AI companies possess the resources to do so. Furthermore, Anthropic continues to spend billions on Nvidia compute. Brad Gastwirth, global head of research and market intelligence at Circular Technology, argued that this does not represent a major near-term threat, as having more chip options does not erase Nvidia's advantages in performance, networking, and software.

Fuentes:Business Insider

Investigación

Chinese Robotics Firm RoboParty Releases Fully Open-Source Humanoid Robot RP1

Tema · RoboPartyMaterial
2026-10-27 20:57 GMT+8

Chinese humanoid robot startup RoboParty recently released its latest product, RP1, claiming it is "fully open-source." This move means developers can not only call high-level APIs but also delve into underlying hardware and software details, including mechanical drawings, electrical systems, proprietary joint modules, and motion control algorithms.

In previous robotics open-source projects, vendors often only opened partial SDKs or algorithm code, making it difficult for researchers to locate and modify issues related to motor response, low-level control, or mechanical structures, creating so-called "technical black boxes." RoboParty's complete transparency of the R&D chain aims to solve widespread problems in research institutions, embodied intelligence companies, and university education regarding hardware adaptation difficulties and limited research autonomy.

RP1 stands 135cm tall and offers various body configurations with 24 to 30 degrees of freedom, demonstrating real-world movement capabilities such as continuous walking, human motion tracking, disturbance recovery, and getting up after falling. Its core hardware is RoboParty's self-developed ROMOMO integrated rotary joint, with a nominal peak torque range of 25–160 N·m, using an inner rotor motor paired with an NGW planetary reducer. Additionally, RP1 has released URDF, MJCF, and mesh model resources, along with the hhtools human motion retargeting tool, supporting simulation training workflows.

Although the scope of open-sourcing has expanded, more fundamental engineering files like CAD and BOM remain partially reserved, with specific availability determined by the project repository. RoboParty has opened beta applications for RP1, attempting to promote shared foundational resources in the humanoid robotics industry by building a three-part system comprising Open Body, PartyOS, and Developer Ecosystem.

Fuentes:量子位

Investigación

OpenAI Announces Three Scientific Breakthroughs with 'Shepherd' Model, Faces Criticism Over Gatekeeping

Tema · OpenAIMaterial
2026-10-28 04:48 GMT+8

US-based AI company OpenAI announced on October 26, 2026, three significant breakthroughs in mathematics, computer science, and physics, introducing a new release mechanism where external experts act as "shepherds" to organize and interpret the AI-generated raw research data. However, this approach has drawn sharp criticism from the academic community, with several prominent scholars accusing OpenAI of monopolizing early access by handpicking specific experts, thereby violating principles of open scientific transparency.

According to The Verge, the three results concern the local smoothing conjecture in harmonic analysis (resolving the Kakeya conjecture), the refutation of the Exponential Time Hypothesis (ETH) in computational complexity, and an existence proof for Quantum Chromodynamics (QCD). For the mathematical result, OpenAI directly released the AI output and Lean formalization; for the computer science finding, it formed a collaboration group led by experts including Russell Impagliazzo, committing to publicize all raw data; for the physics result, OpenAI deferred the release timing entirely to the external expert group.

While some experts, such as Brown University professor Brendan Hassett, noted that the mathematical claim "looks well supported," many researchers expressed dissatisfaction with the tiered disclosure method. Constantin Kogler, a researcher at the Institute for Advanced Study, stated, "It doesn't make any sense that first only a few distinguished scientists can look at them." Andreas Karch, a physics professor at the University of Texas at Austin, was more blunt: "The fact that they handpick the group is a disaster... The large labs have enough power as it is. They should not be picking winners and losers."

Just a day before the announcement, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), which OpenAI had previously pledged to consult, issued a statement opposing this practice. AGMAI recommended that AI companies release outputs directly on scholarly repositories and provide funding to encourage broad participation, emphasizing that understanding should develop in the open with equal access for all. OpenAI responded that it is experimenting with different strategies to balance information asymmetry with community engagement and welcomes feedback to improve the process.

Fuentes:The Verge

Investigación

Seagate Reports Fiscal Q1 2027 Results: $4.2 Billion Revenue, Record Profitability and Free Cash Flow

Tema · 希捷HAMR+SMR硬盘Material
Verificado 2026-10-28 04:51 GMT+8

Seagate Technology Holdings plc (NASDAQ: STX) reported its financial results for the fiscal first quarter ended October 2, 2026, on October 28, 2026. The quarter saw revenue of $4.2 billion, a GAAP gross margin of 57.3%, and a non-GAAP gross margin of 57.8%. GAAP diluted earnings per share (EPS) was $7.59, while non-GAAP diluted EPS was $7.68. Cash flow from operations reached $1.7 billion, and free cash flow was $1.5 billion.

Dave Mosley, Seagate's chair and chief executive officer, stated that the company delivered another quarter of record profitability and its highest ever free cash flow in the September quarter, demonstrating the earnings power of the business amid a strengthening demand environment. Mosley noted that data creation, retention, and reuse continue to compound storage demand, with long-term customer demand visibility reinforcing the view of a multi-year growth phase. The HAMR-based technology roadmap enables the company to address growing exabyte demand through areal density gains while maintaining supply discipline.

Regarding capital allocation, Seagate retired approximately $1.2 billion in principal debt during the fiscal first quarter, lowering gross debt to $2.4 billion. The company returned $568 million to shareholders through dividends and share repurchases and increased its quarterly cash dividend by approximately 3% to $0.76 per share. As of the end of the quarter, cash and cash equivalents totaled $1.3 billion, with 230 million ordinary shares issued and outstanding. Subsequent to the quarter-end, the company retired an additional $750 million of principal debt, bringing gross debt down to $1.6 billion.

For the fiscal second quarter of 2027, Seagate provided the following business outlook: revenue of $4.8 billion, plus or minus $100 million, and non-GAAP diluted EPS of $9.20, plus or minus $0.20. This guidance excludes known pre-tax charges related to estimated share-based compensation expenses.

Fuentes:s24.q4cdn.com

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

Carlson’s argument has a boundary that must be acknowledged: investors in the Theranos case were completely unfamiliar with blood testing technology; their errors may stem more from domain ignorance and celebrity herd behavior than from intelligence itself. The studies by West et al. measured general cognitive biases, not specifically investment decision-making scenarios. However, the direction of the mechanism still holds: when a person attributes past success entirely to their own abilities, they will overestimate their judgment in any new field. The higher the intelligence, the smoother this self-persuasion becomes and the harder it is to interrupt with external questioning.