LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-0499 posts

Atualizar

Smart TV OS Maker Titan OS Raises $46 Million, Grows Users to 30 Million and Adds New Hardware Partners

Tópico · Titan OS融资扩张Material
2026-11-04 23:00 GMT+8

Titan OS, a Barcelona-based smart TV operating system manufacturer, has announced a new $46 million funding round. The capital comes from debt facilities provided by funds and accounts managed by BlackRock, along with existing investors Highland Europe and Mangrove Capital Partners. Combined with the $58 million secured last year, the company's total funding now stands at $110 million.

Titan OS promises TV makers better value from unit sales by loading an operating system that generates money through partnerships with FAST channels (free ad-supported television) and ads on different screens. The company currently serves 30 million users, up from 18 million last year. This includes being a sales partner for TiVo in some regions and Tubi in the UK.

In terms of hardware partnerships, Titan OS has added Sharp, AOC, Vestel, and Skyworth this year. Notably, Dutch electronics maker Philips has switched to Titan OS and ditched Google TV for its 2026 models, marking a significant success for the startup's strategy.

With AI driving up chip costs and hardware sales slowing, TV manufacturers are increasingly reliant on partners like Titan OS that share revenue after the sale. Hyoungung Kim, the newly appointed CFO, stated that the new funds will focus on content discovery optimization and using AI to improve ad recommendation and segmentation, thereby expanding advertiser footprint and audience engagement metrics.

Fontes:TechCrunch

Pesquisa

Canadian AI Firm Cohere Launches Embed 5 and Rerank 5 Retrieval Models on OpenRouter

Tópico · CohereMaterial
2026-11-04 23:00 GMT+8

Canadian artificial intelligence company Cohere has announced that its latest retrieval models, Embed 5 and Rerank 5, are now available on the OpenRouter platform. This integration allows developers to access these models through a single API, using their existing OpenRouter API keys and credits without the need for additional SDKs or complex configuration.

Embed 5 and Rerank 5 represent Cohere's most powerful retrieval tools, designed to handle enterprise documents and complex retrieval tasks. The announcement highlights a key architectural feature: Embed 5 Pro and Embed 5 Fast share the same embedding space. This means developers can build indexes using the high-quality Embed 5 Pro while querying with the low-latency Embed 5 Fast, significantly reducing response times while maintaining retrieval accuracy. According to Cohere's internal tests across 40 datasets, this hybrid strategy retains 98.4% of the retrieval quality of an all-Pro approach, compared to 96.6% if both sides use the Fast model.

In addition to the embedding models, Rerank 5 provides calibrated relevance scores, making threshold filtering more consistent across different queries and domains. This helps eliminate weakly relevant context before it reaches the large language model. For developers already using OpenRouter, this update simplifies the process of building Retrieval-Augmented Generation (RAG) pipelines, allowing for greater flexibility in balancing cost, speed, and accuracy.

Fontes:Cohere

Pesquisa

a16z Partner Jen Kha Argues the 'Safe' Label for Tech Buyout Funds Is Failing

Tópico · 科技并购基金表现争议Estrutural
2026-11-04 23:01 GMT+8

In November 2026, a16z partner Jen Kha published a long-form article weaving a real estate analogy throughout its argument. Her core thesis is that tech buyout funds are not merely facing cyclical headwinds, but that the structural conditions they rely on for survival have vanished. She characterizes tech buyouts as a 'late-cycle' asset class—most effective only after a technology wave has settled and business models are predictable enough to serve as collateral for borrowing. When a new technology wave arrives, all prior assumptions become invalid, which is precisely the environment where venture capital excels.

She employs a consistent analogy: historically, tech buyouts were the 'good house on an ideal street with a steady stream of buyers' in the private markets; venture capital was the 'undeveloped lot on the other side of town.' The problem arises if you bought at the peak (she explicitly names 2021), hold an adjustable-rate mortgage, and the entire neighborhood is being rezoned—the lender will call.

The data supporting this judgment is quite specific. Over the past decade, US buyout funds returned an annualized 14.5%, underperforming the Russell 3000 index, while US venture capital returned an annualized 17.6% over the same period, outperforming the Russell 3000 by approximately 2.3 percentage points—and this was achieved without leverage. More critically, comparing top-tier funds: for vintages raised between 2007 and 2016, the top 5% of US venture funds had a DPI (Distributed to Paid-In) of 5.3x, whereas the top 5% of buyout funds reached only 3.5x. In terms of TVPI (Total Value to Paid-In), top venture funds achieved roughly 2.2 times the TVPI of top buyout funds.

Jen Kha further notes that this gap is not accidental. In venture capital, winners tend to keep winning—the best companies typically call the best VC firms first, meaning a venture firm’s track record remains a significant predictor of its next fund. In contrast, since 2000, a buyout fund’s top-quartile performance in one vintage has not reliably predicted the next. This implies that 'picking the right VC' is easier than 'picking the right buyout manager.'

She acknowledges that the 2010s were the golden era for buyouts: low interest rates drove steadily rising multiples, and Bain data indicates this contributed to over 50% of buyout returns. Global buyout assets under management ballooned from $1.6 trillion in 2015 to $3.9 trillion in 2021. But those conditions—falling rates, multiple expansion, and a large supply of 'fixer-upper' family businesses—are no longer present. Today, PE mostly buys companies already owned by another sponsor or takes public companies private, usually with cost-cutting plans. In 2025, approximately half of US PE exits by count were sales to another sponsor.

The sharpest criticism targets the valuation mechanism itself. Venture valuations are marked by the external market: a company’s valuation rises only when it completes a new funding round and more investors bring larger checks. Buyouts are the opposite—the fund owns the company outright, and the valuation is decided behind closed doors by the holders. Even if the business is struggling, it is easy to bury your head in the sand and maintain the original valuation. She quotes industry wisdom: 'There is more fiction written in Excel than in Word.'

Fontes:a16z.news

Pesquisa

OpenAI Makes Daybreak Red Cybersecurity Model Available in AWS GovCloud (US-East)

Material
2026-11-04 15:00 GMT+8

On November 4, 2026, OpenAI announced that GPT-5.6 Cyber, a model purpose-trained for cybersecurity, is now available to approved customers in the AWS GovCloud (US-East) region through Daybreak Red. This expansion builds on the previous availability of Daybreak models on Amazon Bedrock, aiming to allow government security teams to apply advanced cybersecurity capabilities within the AWS environments they use to support critical missions.

Daybreak Red provides authorized security practitioners with tools for vulnerability research, exploit validation, and security testing. GPT-5.6 Cyber can help defenders investigate complex software weaknesses, reproduce vulnerabilities in controlled settings, and develop mitigations. OpenAI notes that qualified security professionals should review findings and proposed fixes before implementation.

For government organizations, adopting such frontier capabilities requires more than just model performance; teams must also operate within established security, governance, and operational requirements. Availability in AWS GovCloud (US-East) gives eligible teams a path to incorporate GPT-5.6 Cyber into their workflows through Amazon Bedrock, helping connect vulnerability research to practical improvements in defense. The service offers verified defenders greater latitude for authorized cybersecurity work, with fewer refusals for these tasks alongside stronger identity verification, monitoring, and access controls.

Notably, inference data is not used to train models, and customers do not need to share their Daybreak Red data with OpenAI. Access requires both OpenAI Daybreak Red approval and AWS provisioning.

Fontes:OpenAI

Pesquisa

OpenAI Announces Acquisition of Pierre Computer Company to Integrate Open-Source Git Tools into Codex

Material
2026-11-04 22:30 GMT+8

U.S.-based artificial intelligence company OpenAI announced on November 4, 2026, its intention to acquire Pierre Computer Company, Inc. Pierre is a developer of open-source tools, notably Code Storage and the widely used Trees and Diffs libraries. OpenAI stated that these tools are currently integrated into its Codex application to assist developers in navigating files and reviewing code changes.

The strategic rationale behind the acquisition is to strengthen the infrastructure capabilities of Codex as a platform for AI coding agents. OpenAI noted that traditional systems surrounding Git need to evolve as more software development work occurs in parallel via AI agents. By integrating Pierre’s technical expertise in code navigation, diff visualization, and syntax highlighting, OpenAI aims to set standards for modern software development in the 'agentic era,' positioning Codex as the primary environment for developers and AI agents to collaborate on complex projects.

According to the announcement, Pierre’s open-source projects will remain open-source after the transaction closes, with OpenAI committing to continue contributing to them. Existing Code Storage customers can continue using the service during the transition period until its features are fully migrated into Codex. Notably, the acquisition remains subject to customary closing conditions, including regulatory approvals; both companies will operate independently until the deal is finalized.

Fontes:OpenAI

Pesquisa

SemiAnalysis: SpaceX Greenfield Data Center Could Be Built in Five Months, New Target Compressed to Four

Tópico · SpaceX数据中心建设Material
2026-11-04 04:44 GMT+8

In a paid analysis released in November 2026, SemiAnalysis put forward a core judgment: SpaceX may be becoming the world's fastest data center builder, even outpacing its Chinese counterparts. This judgment is supported not by a single project, but by a set of cross-verified data points—the retrofit of Colossus 1 took 122 days to complete 150 MW, the retrofit of Colossus 2 was shortened to 91 days for 210 MW, and Project SPAM, which broke ground on October 5, 2026, targets bringing its first batch of computing power online in December, potentially breaking its own record again. Regarding greenfield projects, Memphis's MiniHard has already been completed in about five months, while the newly purchased "Gun Range" plot, located just south of the Tennessee state line, is expected to take only about four months. The industry typically considers an eight-to-twelve-month construction period as reasonable for greenfield data centers; SpaceX has compressed this by more than half.

Where does this speed come from? SemiAnalysis identifies three layers of mechanisms. The first is campus reuse. Starbase, SpaceX's rocket manufacturing and launch base in Texas, is traditionally used for spacecraft assembly and launch preparation. However, from clearing the ground at the end of August to entering full vertical construction by the end of September, the campus now simultaneously hosts two data centers, and SemiAnalysis judges their scale to possibly exceed one gigawatt. TeraFab—Elon Musk's planned 100-million-square-foot, 10+ GW chip factory project—actually initially started as a data center before pivoting to chip manufacturing. SemiAnalysis believes it is now likely rejoining data center functions, potentially offering capacity for third-party rental. The second layer is Tesla's modular MEP equipment manufacturing. SemiAnalysis believes Tesla is producing prefabricated MEP equipment for SpaceX, bypassing traditional supply chain delivery bottlenecks. This is analogous to Crusoe self-producing switchgear, but on a much larger scale. The third layer is the flexibility of the power strategy. Instead of waiting for traditional grid interconnection approvals, SpaceX uses a combination of truck-mounted gas turbines, large-scale energy storage, and later grid integration. SemiAnalysis tracks over 8 GW of behind-the-meter capacity, with Colossus 2 currently deploying an energy storage system nearing 6 GWh, which could become the largest single-site energy storage deployment in the United States.

The power issue is not without gaps. SemiAnalysis points out that after signing binding agreements for the Southaven power plant at Colossus 2 in July 2026, the legally operable nameplate capacity will drop from 2.7 GW to 1.24 GW by July 2027. Subtracting the approximately 1.2 GW of grid connection leaves a gap of about 1.3 GW between computing capacity and 1.24 GW of nameplate power generation. Considering efficiency derating and total facility load, the shortfall could reach 300 MW. SemiAnalysis believes this gap can be bridged in three ways: ultra-low PUE design (approximately 1.10 to 1.15), differentiated power demands for inference versus training clusters, and orchestrated scheduling of storage and gas—where gas turbine cycles charge the storage, and storage supplements baseload during peak periods. Additionally, some clusters require high availability for customers, while others are used for internal training and can tolerate fluctuations; SLA tiering itself can release total capacity.

Fontes:newsletter.semianalysis.com

Pesquisa

Berlin Fintech Pliant Raises $80M Series C from Seaya Nova and Speedinvest to Build AI Agent Payment Infrastructure

Tópico · PliantMaterial
2026-11-04 22:00 GMT+8

Berlin-based B2B payments platform Pliant has closed an $80 million Series C funding round led by Seaya Nova and Speedinvest, according to a report by Tech Funding News. The oversubscribed round also saw participation from existing investors Illuminate Financial, Motive Partners, and Molten Ventures. The new capital will support Pliant's expansion into the US and Canada, entry into additional international markets, and the development of what the company calls "agentic payments."

Founded in 2020 by Malte Rau and Fabian Terner, Pliant provides commercial card services, payment accounts, and spend management. The company states that its annual recurring revenue (ARR) has now surpassed the $100 million mark. This round doubles the size of Pliant's $40 million Series B completed in April 2025. Based on the reporter's calculations, Pliant's total equity funding now exceeds $180 million.

Pliant's strategic focus is integrating AI software agents with its licensed payment infrastructure. The company argues that as external assistants like Claude or ChatGPT, as well as in-house agents, increasingly handle corporate finance tasks, they require connections to regulated payment systems via open standards such as the Model Context Protocol (MCP), rather than just API integrations. As a principal member of Visa and Mastercard operating under a European e-money license, Pliant positions itself as the compliant backend for these autonomous transactions.

This move highlights the growing intersection of AI and regulated financial services. Grand View Research estimates the agentic commerce market at $5.71 billion in 2025, projected to reach $65.47 billion by 2033. However, Pliant faces competition from larger, better-funded players like Ramp (valued at $44 billion) and Brex (acquired by Capital One). While AI agent payments are still in their early stages, Pliant aims to capture the market by providing the necessary licensed infrastructure for this emerging workflow.

Fontes:Tech Funding News

Pesquisa

AI Startup Oumi Extends Frontier AI Engineer to Build Agents and Retrain Underlying Models

Tópico · Oumi FAiE平台Material
2026-11-04 22:00 GMT+8

Seattle-based enterprise artificial intelligence platform company Oumi PBC announced on November 4, 2026, that it has extended the capabilities of its “Frontier AI Engineer” (FAiE) agent. The update enables the system to not only build AI agents for customers but also retrain the underlying models those agents rely on. This move targets a common problem surfacing when agents leave the demo stage: failing agents often stem from bad prompts, broken tools, or weak base models, traditionally requiring skills found at frontier AI labs to fix.

According to SiliconANGLE, Oumi CEO Manos Koukoumidis stated that better prompts and larger generic models will only take agents so far. He argued that successful agents improve “by training the weights of the models powering them, learning from their own work.” The new FAiE allows teams to describe goals in plain language; the system then defines necessary tools and environments, converting production failures into evaluation sets and training data for subsequent rounds.

Technically, Oumi has added Direct Preference Optimization (DPO) to its existing supervised fine-tuning, enabling models to train on examples of user-preferred behaviors. It also introduced Group Relative Policy Optimization (GRPO), which grades models on task outcomes. These methods can be chained in single multistage training runs. Additionally, the new “Bring Your Own Harness” feature lets customers keep their existing orchestration, tools, and memory while connecting to Oumi for evaluation and training, ensuring models are tuned within the same environment where the agent operates daily.

This initiative reinforces Oumi’s business argument regarding “model ownership”: most enterprise agents run on general-purpose frontier models not optimized for specific customer workflows, whereas a specialized model owned outright by the customer may perform better at lower cost. GRPO is currently launching in private preview, and the company is accepting access requests.

Fontes:SiliconANGLE

Pesquisa

UK Fintech Revolut Valued at $115 Billion After Secondary Share Sale Ahead of US Bank Launch

Tópico · Revolut上市与扩张Material
2026-11-04 22:10 GMT+8

Investors in UK-based fintech platform Revolut have completed a secondary share sale that values the company at $115 billion, 53% higher than the valuation established in another share sale a year ago. The transaction was led by firms including Glade Brook Capital Partners, Coatue, Durable Capital Partners, T. Rowe Price Associates, and Baillie Gifford, but did not bring new capital into Revolut. The company declined to disclose how much stock was sold.

This valuation jump comes as Revolut prepares to enter the US market. The company has received conditional approval from the U.S. Office of the Comptroller of the Currency (OCC) to establish a national bank, which it plans to launch in the first half of 2027, subject to final regulatory approval. Founded in 2015, Revolut has grown from a money-transfer app into a financial services platform with more than 80 million retail customers across over 40 countries. This year, it launched banks in the UK, Mexico, Australia, and New Zealand, and secured licenses in France, Colombia, and Argentina.

Financial data indicates significantly improved profitability for Revolut. In 2025, the company achieved profit margins of 38% on revenue of $6 billion. While its core business targets consumers, its enterprise banking division, Revolut Business, has become an increasingly lucrative source of income. James Gibson, head of Revolut Business, stated that the division accounts for about 20% of the company's gross profit, processing more than £36 billion ($48.6 billion) in transactions monthly, including over £1 billion in the US alone. The division generates more than £950 million in annualized revenue and holds over £18 billion in customer deposits.

As the valuation rises, attention turns to Revolut's initial public offering (IPO) plans. Founder Nik Storonsky has considered a dual listing in New York and London, indicating a preference for a primary US listing. Reports suggest an IPO could value Revolut at up to $200 billion, but this would likely occur no earlier than 2028, depending on market conditions.

Fontes:TechCrunch

Pesquisa

CVS Health Confirms Split of Pharmacy and Insurance Businesses, Reversing 2018 Aetna Merger

Tópico · CVS Health拆分Material
2026-11-04 22:05 GMT+8

U.S. healthcare giant CVS Health confirmed on Tuesday night that it will split its retail pharmacy business from its insurance operations into two separate companies. This decision reverses the 2018 merger that combined insurance behemoth Aetna with the 9,000-store retail pharmacy chain to create a $117 billion empire.

"We believe based off of where we are and how we’re performing, the businesses are best served by splitting," CVS Chief Executive David Joyner told The Wall Street Journal.

The split comes amid increasing regulatory scrutiny and political pressure on the U.S. healthcare industry. Critics had argued that the 2018 merger led to less competition and higher prices for consumers. With Democrats expected to regain control of Congress in January, pressure on large healthcare conglomerates is anticipated to grow further. By splitting, CVS may aim to reduce antitrust risks and allow the two new entities to operate more flexibly in their respective markets.

Fontes:semafor.com

Pesquisa

Zvi Mowshowitz: OpenAI's New Model Spec Examples Are Too Simple to Teach Models How to Handle Truly Difficult Cases

Tópico · OpenAI模型规范争议Material
2026-11-04 00:48 GMT+8

OpenAI’s Model Specification (Model Spec) is potentially its most important document, outlining what the company aspires for the model to do and how it should think. In early November 2026, AI commentator Zvi Mowshowitz published an analysis article comparing the new and old versions line by line. His core criticism was not directional—he acknowledged that the new specification’s overall philosophy is similar to the old one and that many changes are improvements—but pedagogical: a large number of newly added behavioral examples suffer from 'overdetermined answers,' failing to teach the model how to judge truly difficult cases.

Mowshowitz refers to this problem as 'Goldilocks' answers: two extremes are marked BAD, while the middle ground is marked GOOD. However, neither the model nor the reader learns how to choose between two 'GOOD' options or how to weigh trade-offs in genuinely ambiguous situations. He wrote: 'If we are producing a deontological document, then we need difficult cases that make good law, showing where we would not otherwise know what to do. Easy cases do not make bad law, but neither do they make good law.'

He illustrated this issue with several specific examples. In the section on context usage, the specification provided an example of a user with a peanut allergy ordering Thai food: The GOOD response was for the model to notice the allergy information in the order history and proactively ask; the two BAD responses were ignoring the information and directly adding an allergy note. Mowshowitz agreed with the general direction but pointed out that the example itself was poorly chosen—Pad Thai comes with crushed peanuts by default, and someone with a severe peanut allergy would typically not order this dish, let alone ask an AI to randomly find a restaurant to place an order. When the premise conditions of an example are unnatural, the rules extracted by the model may lead to unexpected generalizations.

In the 'Do Not Lie' section, the specification provided an example of helping a user find a breeder for Scottish Fold cats: The GOOD response was to provide breeder information while mentioning the joint diseases the breed may cause; the BAD responses were omitting the mention entirely and blocking the user first before answering. Mowshowitz’s follow-up question was: What principle is this example actually teaching? Is it because of concern for animal welfare? Because countries like the Netherlands and Austria have legal bans? Because veterinary costs are high? He attempted to test similar scenarios in actual conversations and found that the model proactively warned about Scottish Folds but did not warn about Pit Bulls (which can be dangerous); when questioned, the model admitted the inconsistency. He concluded that principles need to be explicitly stated, rather than contained within an 'overloaded' example.

In the 'Forced to Pick a Side' section, the specification provided three examples regarding whether Franklin D. Roosevelt or Ronald Reagan was better, with all GOOD responses choosing FDR. Mowshowitz pointed out that having three choices with the same political leaning on a question without an objective answer would lead the reading AI and technical staff to draw unintended inferences. His more fundamental criticism is that when you provide examples in a root-level document, AIs and technical staff will attempt to extrapolate from them, including potentially drawing unintended conclusions. The details of the examples matter, especially when the principles themselves are not explained by other text.

Fontes:thezvi.substack.com

Pesquisa

xAI Powers Colossus 2 with Repurposed Jet Engines to Bypass US Grid Delays

Material
2026-11-04 20:57 GMT+8

US technology companies are increasingly turning to engines from retired aircraft to power new AI data centers as quickly as possible, bypassing time-consuming grid connections, according to a report by German tech outlet heise online citing The New York Times.

The report states that xAI, the artificial intelligence subsidiary of Elon Musk's SpaceX, has installed eight converted aircraft engines in Mississippi to supply power to its Colossus 2 data center near Memphis, Tennessee. The New York Times confirmed this situation through official documents and satellite imagery.

This 'creative' solution stems from growing power bottlenecks in the expansion of US AI infrastructure. While these modified engines are less efficient than standard gas turbines and more environmentally harmful, they significantly reduce deployment times. According to ProEnergy, a company specializing in such conversions, each Boeing 747 turbine can deliver approximately 50 megawatts of power.

A January report by the US Energy Information Administration (EIA) noted that engines from retired military aircraft at the Davis-Monthan Air Force Base in Arizona could theoretically generate up to 40,000 megawatts, exceeding the total output of all power plants in that state. As demand for old aircraft components surges, aircraft graveyards are becoming new resource hotspots.

Fontes:heise online

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

This valuation smoothing is maintained through several mechanisms. When no outside buyer appears, more sponsors sell companies to new continuation funds they also manage, funded by secondary investors. These 'continuation vehicles' accounted for 14% of PE exit value in the first half of 2026, up from just 5% in 2020. Some funds go even further, borrowing against their entire portfolio to distribute cash to LPs, causing distribution figures to rise without any company actually being sold—these NAV loans add a second layer of leverage on top of existing corporate debt.

The 2021–2022 vintages are where problems are most concentrated. Many companies can no longer pay interest in cash, so they capitalize it (PIK). For these borrowers, the loan-to-value ratio jumped from 44% to 75% in a single year. Of the debt defaults occurring in 2025–2026, 70% originated from 2021 or 2022 deals. Thoma Bravo acquired Medallia in 2021 for approximately $6.4 billion; this year, it handed the company over to a group of lenders led by Blackstone, wiping out about $5.1 billion in equity value. If a sponsor of that caliber cannot make a 2021 deal work, the issue lies with the vintage year itself.

Jen Kha concludes that private equity will not be 'cooked' overnight; capital is sticky, and the best buyout managers will adapt. However, the conditions that made the 2010s a golden era—a supply of fixer-uppers, continuously falling rates, and funds buying from each other at higher prices—will not return. A few rate cuts may alleviate debt pressure, but they cannot rebuild that era. She advises LPs to contact managers today to ask how much interest is being added to loans for the 2021–2022 vintages, and then seek an exit path.

Finally, she points out that if intuition suggests trimming buyout exposure in favor of private credit or secondaries, one must look at the other side of that trade-off: private credit loans to SaaS companies grew from approximately $8 billion in 2015 to over $500 billion by the end of 2025, but the lender’s upside is capped at the interest owed; secondaries primarily buy existing buyout stakes at a discount. Either way, you still hold the same buyout companies, just with less potential gain.

Boundaries that must be retained: SemiAnalysis has identified a capacity path exceeding 9 GW, but the full 13 GW has not yet been confirmed. The judgment on construction speed is based on cross-referencing satellite imagery, local filing documents, and supply chain information, representing proprietary model outputs from a research institution rather than independently verified facts. The 13 GW figure itself comes from Elon Musk's statement at SpaceX's first earnings call; SemiAnalysis deems it feasible but acknowledges it depends on whether power approvals, equipment deliveries, and construction execution can continue to match. The purpose of the TeraFab data center—whether for self-use or third-party rental—has also not been officially confirmed by SpaceX. SemiAnalysis only infers from recordings of meetings with local officials that representatives did not rule out the possibility of third-party leasing.

For AI industry investors, the practical significance of this analysis lies in repricing a variable: if a company can complete greenfield construction in five months and retrofits in 91 days, then the bottleneck for AI computing supply will no longer be construction cycles, but chip availability and power approvals. This compresses the time from capital expenditure to revenue recognition and means that ecosystems possessing vertical integration capabilities (Tesla manufacturing, SpaceX construction, xAI usage) hold a structural speed advantage over cloud providers relying on external contractors and traditional grid interconnections. SemiAnalysis connects this judgment with its previous analysis that contracts with Anthropic and Google have turned SpaceX into the most profitable neocloud provider. It believes that if the 13 GW target is achieved, the corresponding annual recurring revenue will exceed $300 billion.

Mowshowitz does not oppose all changes in the specification. He holds a positive view of the new discussion on 'letter and spirit,' the emphasis on corrigibility, and the expansion of the impact function (treating budget as a cap rather than a target). His criticism focuses on a methodological issue: a deontological specification document should select examples that humans would not know how to handle without guidance, rather than those where intuition already provides a clear answer. The former makes good law; the latter only confirms common sense.

The boundaries of applicability for this criticism are worth noting. Mowshowitz’s argument is based on the assumption that examples in the model specification are indeed used for model training or behavior shaping, and are not just references for human readers. If the primary function of the examples is to convey corporate intent to the public and developers, then simple and clear cases might be exactly the right choice. Additionally, he himself admits 'this is not a terrible position'—the old specification was already a solid document with many strong details, and the new version is philosophically and practically similar to the old one. His criticism is more like a suggestion for improvement direction rather than an accusation of fundamental flaws.