LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-10-29198 publicaciones

Actualizar

Construction Physics author Brian Potter: Fusion could significantly reduce the cost of isotope production

Tema · 同位素制造成本Material
2026-10-29 23:45 GMT+8

Technetium-99m, widely used in medical imaging, relies on the decay of molybdenum-99, which is produced by only a handful of nuclear reactors worldwide. In 2024, maintenance shutdowns at some of these facilities triggered global shortages. Priced by the gram, isotopes are among the most expensive materials on Earth, and their supply chains are extremely fragile. Brian Potter, author of Construction Physics, analyzed the technical roots of this dilemma and proposed a potential solution: fusion.

One of the primary methods for producing isotopes today is using particle accelerators. These devices use electromagnetic fields to accelerate charged particles, meaning they can only interact with electrons, protons, or positively charged atomic nuclei. The Coulomb repulsion between protons and target nuclei results in very low conversion efficiency: for example, General Electric Healthcare’s PETtrace 800 uses protons to bombard oxygen-18 to produce fluorine-18, but only about one in every 800 protons successfully enters an oxygen nucleus. A single medical accelerator costs millions of dollars, operates at approximately 75 kilowatts, and produces material roughly the size of a speck of dust per hour.

Another path involves nuclear fission. Fission reactors release neutrons that can bombard target materials to produce isotopes, such as placing cobalt-59 in a reactor to generate cobalt-60 via neutron capture. However, fission neutrons generally have low energy—averaging about 1.5 to 2 million electron volts (MeV)—and in power reactors, moderators further slow them down to about 0.025 eV to sustain the chain reaction. Even without moderation, most fission neutrons lack the energy required to drive many useful nuclear reactions. Additionally, fission reactors are heavily regulated, making them difficult and expensive to build.

Potter points out that fusion reactions reliably produce high-energy neutrons. Deuterium-tritium fusion releases neutrons with energies exceeding 14 MeV, whereas only about 0.1% of fission neutrons exceed 10 MeV. Crucially, fusion facilities can output energy alongside isotope production. While accelerators only produce isotopes—requiring all facility costs to be recovered through isotope sales—fusion facilities, if configured with appropriate reaction combinations, can use revenue from electricity sales to offset construction and operational costs. Potter believes that under suitable arrangements, the production costs for certain isotopes could drop significantly, though he explicitly limits this conclusion to apply at least to specific isotopes and reactions.

A significant reduction in isotope costs would directly impact several sectors. In medicine, actinium-225 is a promising isotope for cancer treatment, but current supplies are extremely limited, allowing only a small number of patients to be treated annually; cheaper manufacturing processes could expand access and even enable new medical isotopes. For nuclear fuel, fusion reactors require lithium-6 to breed tritium within the cycle, fission reactors need highly pure lithium-7, and some new reactor designs, such as TerraPower’s molten salt reactor in the United States, require chlorine-37. Currently, these isotopes are either produced only in gram quantities or rely on expensive enrichment processes in Russia and China. In materials science, isotopically pure silicon-28 is considered valuable for building quantum computers due to its distinct quantum mechanical properties, and isotopically pure diamond has far higher thermal conductivity than standard diamond. Historically, producing isotopically pure materials has been so difficult that many of their properties remain unknown.

Fuentes:construction-physics.com

Investigación

Jake Tennant of ElevenLabs on Product and Growth Directions: Removing Fields Improves Retrieval Agent Accuracy More Than Prompt Tuning

Tema · ElevenLabsMaterial
2026-10-29 20:00 GMT+8

Jake Tennant, who works on product and growth at ElevenLabs, introduced the company's operational experience with its internal ElevenAgents in a blog post. These agents cover sales, support, engineering, and product marketing teams; the company reports saving over 1,200 hours monthly and answering more than 5,000 questions while creating over 150 tickets.

However, what truly warrants attention from peers is not these numbers, but the engineering judgment provided by the author in the 'What We Learned' section. Tennant wrote that the greatest improvement in answer quality for their retrieval agents came from removing fields rather than adjusting system prompts. Every extra column or object field gave the model another opportunity to guess incorrectly; trimming the data schema down to a critical subset improved accuracy more than any prompt changes they had tried.

This finding creates tension with current mainstream prompt engineering approaches. Many teams default to fine-tuning at the prompt level, but Tennant's experience suggests that when agents rely on external data retrieval, the data structure fed into the model itself becomes a larger quality bottleneck. Fewer fields mean less room for the model to guess.

The author provides a specific implementation example: fields such as deal stage or account type in Salesforce are picklists, meaning a fixed set of valid values. If left to generate freely, the model might create a plausible-sounding option that does not exist in Salesforce. ElevenLabs' approach is to fetch the list of valid values at startup and inject them into the MCP tool's type definitions, ensuring the model can only choose from real options.

Additionally, they added a task verification step after ticket creation: after the agent creates a Linear ticket, it checks whether the fields are correct and if the ticket was assigned to the right team. Tennant stated this has reduced occasional hallucinations and misassignments.

It should be noted that this is an account from ElevenLabs regarding its own products. The 1,200 hours saved lacks a methodological explanation, and the engineering experience may be tied to their specific tech stack and internal data scale. However, these specific practices—prioritizing field removal over prompt tuning, using type constraints instead of free generation, and verifying after creation—are directly testable for other teams within their own environments.

Fuentes:ElevenLabs

Investigación

NVIDIA Says TF32x3 Accelerates Machine Learning Interatomic Potentials While Preserving Energy Conservation, Single-Pass TF32 Fails to Converge

Resumen rápido
2026-10-29 23:00 GMT+8

Machine learning interatomic potentials (MLIPs) use neural networks to predict atomic energies and forces at a fraction of the cost of quantum mechanical methods like density functional theory, but their precision requirements remain high. Forces are derivatives of energy with respect to atomic positions; in molecular dynamics simulations, forces at every step determine updates to atomic velocities and positions, meaning numerical errors accumulate along the trajectory. Therefore, even though GPUs offer much higher throughput for low-precision matrix multiplication than for FP32, many MLIPs continue to rely on FP32 inference.

In a technical blog post, NVIDIA evaluated two Tensor Core FP32 simulation methods. On Blackwell GPUs, BF16x9 decomposes each FP32 operand into three scaled BF16 components, computes nine pairwise products, and accumulates them in FP32; it can be enabled via the environment variable CUBLAS_EMULATE_SINGLE_PRECISION=1 without modifying model code. On Ampere GPUs, TF32x3 splits each FP32 operand into a leading TF32 component and a residual TF32 component, computing three cross-products (leading × leading, leading × residual, residual × leading) while omitting the smallest residual × residual term. Implemented via CUTLASS's OpMultiplyAddFastF32 operator, inputs and outputs remain in FP32, making the method transparent to the rest of the model.

The real difference lies not in point precision but in physical consistency. The NVIDIA team validated this using three tests: when dragging a CO molecule away from a Pt(111) surface, TF32x3 followed the smooth energy curve of FP32, while single-pass TF32 introduced visible point-to-point fluctuations; in an NVE simulation of a 192-atom water box, TF32x3 showed bounded energy fluctuations comparable to FP32, whereas single-pass TF32 exhibited systematic drift; during structural relaxation of a 539-atom Cantor alloy slab, TF32x3 reached the force convergence standard of 1.0×10⁻³ eV/Å within 218–222 steps (comparable to FP32's 218 steps), while single-pass TF32 failed to converge even after doubling the step limit to 1,000 steps, with forces plateauing around 1.2×10⁻³ eV/Å.

Regarding performance, BF16x9 achieved an end-to-end speedup of 1.7x on a 17,280-atom system (1.1–1.2x on a 192-atom system), with errors falling below the model's own non-determinism levels; TF32x3 accelerated MD steps on A100 GPUs by about 1.5x, with force errors roughly three times the differences between FP32 recompilations. However, the magnitude of acceleration depends on system size and the proportion of total runtime spent on GEMM operations—smaller systems see more limited gains due to higher startup overhead and non-GEMM workloads. TF32x3 performance also relies on kernel tuning specific to matrix shapes and hardware. Neither method requires retraining or changes to the model architecture, allowing them to serve as independent performance layers that can be combined with other software optimizations.

Fuentes:NVIDIA Technical Blog

Investigación

Berkshire Partners Investor Richard Lichtenstein: AI Assistants May Need a Digital Twin

Material
2026-10-29 18:03 GMT+8

Richard Lichtenstein is an investor at Berkshire Partners. In a personal blog post dated October 29, 2026, he proposed an exploratory idea: for AI assistants like Muse, Instinct, and Dots to better handle business tasks autonomously, he believes it is time to build them a digital twin. His exact words were:

"I've been thinking about what would make these tools better, and I think what they really need is a good digital twin."

The tone is one of personal reflection rather than a definitive conclusion. His basic logic is that if AI assistants have a deeper understanding of user preferences and organizational workflows, they can make more decisions with less human intervention.

He borrows the concept of industrial digital twins to explain this approach. BMW has created digital replicas of more than thirty of its plants; before launching new car models, they test on virtual production lines whether components will collide, compressing physical testing time from four weeks to three days. His analogy suggests that if we built a digital replica of a person—a single physical scan enabling consistent sizing for online shopping and pre-hemmed pants based on leg length—return rates would decrease significantly.

However, he acknowledges that this consumer-focused version is impractical, shifting instead to a more feasible approach: creating a mental digital twin for individuals. He gives the example that if he provided an AI with the itineraries of his last five family trips along with frank summaries of what worked and what didn't, he bets it could plan the next trip effectively. Yet he points out this requires considerable upfront investment and may take several trips to pay off. Currently, these assistants lack mechanisms to proactively seek feedback—he booked a restaurant via Instinct, which never asked about his experience, a missed opportunity he identifies.

The article's focus shifts to the business version. Lichtenstein argues that the corporate side may be easier to implement because more ideas are already documented. He envisions a scenario where, with an exact mirror of a company, you could instruct an AI assistant to monitor the top thirty suppliers, automatically developing backup plans and contacting alternative suppliers upon negative news; follow up with delinquent customers while noting special terms and historical exemptions; identify overdue contacts and reinitiate communication.

For private equity investors, he proposes a specific application: upon entering a newly acquired portfolio company, test parts of the value creation plan on the digital twin before executing them in the real business.

However, he provides the most critical caveat in the article: connecting all documents does not equate to building a digital twin. Documents tell you how a company is supposed to work, not necessarily how it actually works. A large amount of important information remains in people's heads. This means that even if a company stores contracts, SOPs, supplier scorecards, and key meeting presentations in SharePoint or Google Drive, and connects AI to ERP, CRM, email, and instant messaging systems, the result is still an incomplete mirror.

Fuentes:artificialinvestment.substack.com

Investigación

No Priors Interview: Serval CEO Jake Stauch on Challenging ServiceNow with AI Coding Agents

Tema · ServalMaterial
2026-10-29 18:00 GMT+8

In the No Priors podcast episode released on October 29, 2026, hosts Sarah Guo and Elad Gil interviewed Jake Stauch, co-founder and CEO of the AI startup Serval. The episode focuses on how Serval aims to transform enterprise IT workflows by introducing code-generation agents, directly challenging ServiceNow's position as the system of record for enterprise automation and service management.

According to the show notes, Stauch explained Serval's core value proposition: enabling enterprise IT users to safely become 'builders.' He mentioned that the Serval platform allows users to leverage AI for development while safeguarding permissions through strict deterministic approvals and an audit trail. Additionally, Stauch discussed the company's ProServal team, which embeds automation engineers directly within customer organizations to provide ongoing support and service.

For readers interested in AI industry investment, this episode offers a window into how AI is penetrating traditional enterprise software markets. Stauch explores AI as the key unlock for taking on ServiceNow, his bet on the continued improvement of coding agents, and why he views consumer AI as a predictor of enterprise trends. He also touches on the idea that the 'SaaS-pocalypse' might look more like consolidation than disruption, and why background agents remain a moonshot in enterprise software.

Fuentes:traffic.megaphone.fm

Investigación

Anthropic IPO Filing: Most Compute Goes to R&D, Not Customer Service; $800 Billion in Future Chip Contracts Revealed

Tema · Anthropic上市Material
2026-10-29 19:30 GMT+8

AI startup Anthropic has filed its S-1 registration statement with the U.S. Securities and Exchange Commission (SEC), revealing detailed financial and operational metrics. According to an analysis by Tom's Hardware, of the $8.9 billion total compute bill in the first half of 2026, Anthropic spent only $2.4 billion directly serving paying customers. The vast majority of the remaining funds were allocated to model training, research, and internal use of Claude. This distribution highlights the current cost structure of frontier AI labs: while inference services have become profitable, the heavy investment required to maintain technological leadership still dominates spending.

The data indicates a significant improvement in revenue efficiency. Revenue per dollar of compute jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company attributes to capacity purchased 12–18 months in advance. After deducting serving costs, Anthropic retains approximately 85 cents for every dollar of revenue, up from just 12 cents in the same period last year. However, this high margin state is partly aided by a compute-constrained environment, where limited supply keeps serving costs muted rather than purely reflecting efficiency gains.

More striking are the massive future capital commitments. The filing details approximately $800 billion in TPU and GPU purchase contracts with suppliers including Broadcom, Google, Amazon, SpaceXAI, Microsoft, Nscale, Lambda, and AMD. Specifically, five-year equipment leases with Broadcom account for about $161.2 billion in rent, contracts with Google for at least $114.5 billion, and deals with Amazon for at least $110 billion for Trainium chips. These multi-year, large-scale fixed or minimum purchase obligations imply that annual cash outflows will surge significantly once deliveries commence.

For investors, the critical tension lies between current high growth expectations and the burden of future infrastructure debt. While Anthropic claims serving costs are currently covered, the estimated average annual cost of its signed future contracts is $103 billion, far exceeding its Q2 2026 annualized revenue pace of about $46 billion. The company warns in the filing that margins may compress as new hardware lands ahead of the revenue it will support. This disclosure provides a concrete sample for assessing the actual difficulty of the AI industry's transition from 'burning cash for growth' to 'sustainable profitability.'

Fuentes:Tom's Hardware

Investigación

Thrive Holdings Agrees to Take ADT Private in $15.3 Billion Deal

Tema · ADT私有化收购案Material
Verificado 2026-10-29 23:36 GMT+8

US home security company ADT Inc. announced today that it has entered into a definitive agreement to be acquired by Thrive Holdings in an all-cash take-private transaction. The deal values ADT at approximately $15.3 billion in enterprise value and about $7.6 billion in equity value. The offer price of $10 per share represents a premium of nearly 46% over ADT's closing price on October 27.

The financing behind the deal adds a layer of intrigue: the acquisition will be funded by funds managed by affiliates of Apollo Global Management. Apollo was previously ADT's owner and sold its remaining stake in the company just five months ago. Thrive Holdings is a permanent-capital investment vehicle formed last year by Josh Kushner's Thrive Capital.

ADT sells security systems, cameras, and smart-home technology to residential and small-business customers, generating recurring revenue through monthly monitoring subscriptions. This marks ADT's second take-private in a decade; Apollo previously took the company private in 2016 before bringing it back to the NYSE in 2018.

The agreement includes a 'go-shop' period running through November 28. The transaction is expected to close in the first quarter of 2027, subject to ADT stockholder approval and regulatory clearances. Citi and Goldman Sachs advised ADT, while Evercore advised Thrive.

Fuentes:Axios

Investigación

German tech media t3n: Iterating an AI agent through everyday language, without a development team

Tema · t3n AI代理工作流Resumen rápido
2026-10-29 22:15 GMT+8

Advertising operations at German tech media outlet t3n involve a repetitive workflow: every closed deal generates a PDF order confirmation listing all services purchased by the client (such as podcast sponsorships or two sponsored articles). Each service requires a separate task, subtask, and deadline to be created in the Asana project management tool. As of mid-August 2026, this work was done manually by the campaign execution team, taking up to 40 minutes per large order.

Instead of relying on development resources, the t3n team asked Claude what materials were needed to automate this process. They provided brief process documentation and 10 real order confirmations, instructing Claude to compare its output against orders the team had already manually entered. From these discrepancies, the first version of the instructions emerged in a single afternoon.

The first version was insufficient for daily use. The team continued working in a Slack thread, pointing out errors in plain language, which Claude incorporated into its own instructions with each correction. After 24 hours, the instructions reached version 12. A key prerequisite was that the toolchain was already in place: Claude was connected to internal company tools, and the Asana board featured fixed fields and templates.

In mid-August, the agent went live with real orders; the team never created a test copy of the board. Since then, it has been sufficient to tag Claude in the Slack thread with the order confirmation. The agent automatically creates the tasks, while the team verifies the output: less than five minutes for simple orders, and up to ten minutes for large ones. The team reported no errors penetrating daily operations during the initial two-plus months.

It is important to emphasize that this is a firsthand account from t3n regarding its own internal processes, constituting an attributed self-report rather than independent verification. The claim of 'two months without error' does not provide the total number of orders as a denominator, making it impossible to calculate the actual error rate. More critically, the entire method relies on the premise that Claude is integrated with internal tools and Asana templates exist—whether this conversational iteration path would yield similar results for teams whose toolchains are not yet connected is not evidenced in the original text.

Fuentes:t3n.de

Investigación

Taiwan's Kaohsiung City Government Unveils Initial Results of 'AI Lighthouse' Project: Integrates Cosmos Reason 2 for Causal Inference, Reduces Pothole Dispatch Time to Approximately 2 Minutes

Tema · 高雄AI灯塔计划Material
2026-10-29 18:54 GMT+8

On October 29, 2026, the Kaohsiung City Government in Taiwan unveiled the initial results of the "Kaohsiung AI Lighthouse" project. Launched in 2024 in collaboration with Nvidia, Chunghwa Telecom, Hon Hai Precision Industry, and other partners, the initiative aims to build a city-scale sovereign AI governance architecture. Its core objective is to shift from a traditional model heavily reliant on manual labor and citizen reports to one where AI acts as the city's brain, proactively detecting anomalies and analyzing their causes through cameras and sensors.

A key technical breakthrough of the project is the evolution of AI capabilities from simple image recognition to understanding and reasoning. The initially deployed Vision Language Model (VLM) can describe urban scenarios in text, such as identifying traffic congestion on specific road segments. Subsequently, the integration of the Nvidia Cosmos Reason 2 (CR2) model added causal inference capabilities, allowing the system to synthesize multiple images and continuous event data to determine the underlying reasons for anomalies, such as distinguishing between traffic accidents, road flooding, or construction blockages. Kaohsiung is also currently testing the next iteration, Cosmos Reason 3.

To ensure AI-generated insights enter actual administrative workflows, the city established the Smart Governance City (SGC) platform. When AI detects issues like potholes, flooding, or environmental pollution, the SGC platform integrates the event information and automatically notifies the relevant authorities. For road inspections, utilizing 45 mobile vehicles and 135 cameras, the Public Works Bureau emphasized that AI-integrated workflows have reduced the time from problem discovery to dispatch to approximately 2 minutes in the fastest cases, compared to tens of minutes or hours previously.

In flood prevention and traffic management, AI has demonstrated early warning capabilities. The Water Resources Bureau integrated 400 water monitoring stations and over 13,000 cameras; during the heavy rain event on August 22, 52% of incidents were proactively detected by AI. The Transportation Bureau reported that using VLM and CR2 models allows for network anomaly alerts 14 minutes earlier than traditional manual records. The Environmental Protection Bureau uses AI to identify black smoke and illegal dumping, pushing alerts within 30 seconds. In a fire case in Gangshan, AI detected the black smoke and notified the SGC platform about 3 minutes before the Fire Department received a report.

Mayor Chen Chi-mai stated that the project resolves the data silos and redundant funding burdens caused by independent systems across different bureaus. He noted that this verified new model of urban governance could be purchased and adopted by other cities in the future, avoiding the costs of rebuilding a city AI brain from scratch.

Fuentes:iThome

Investigación

Gallup Report: US Worker Engagement Hits Lowest Level Since 2011, Remote and Hybrid Workers Hit Hardest by AI

Material
2026-10-29 23:05 GMT+8

According to the latest quarterly survey by consulting firm Gallup, US worker engagement fell to 29% in the third quarter of 2026, marking the lowest level since 2011. This metric tracks employees' focus, motivation, and enthusiasm for their jobs.

The survey revealed that remote and hybrid workers were disproportionately affected. In organizations that have adopted AI tools, fully remote employees saw a five-point drop in engagement, while hybrid employees declined by two points. In contrast, on-site employees and those in roles that cannot be performed remotely saw only about a one-point decrease.

Gallup noted that although earlier research in 2026 indicated a potential thawing in employee attitudes toward AI, this trend has reversed. Workers in "AI-integrated organizations" now report similar overall engagement levels as those in non-AI organizations, but hybrid and remote groups continue to express lower happiness and participation.

The report suggests that physical distance may contribute to psychological distance between employees and their organizations. As companies navigate the use of AI at work, the pace of adoption appears to be outstripping the communication and managerial support needed to make it effective.

Fuentes:TechRadar

Investigación

Oracle Co-CEO: AI Compute Bottleneck Lies in Chip Capacity, Not Power

Tema · 甲骨文芯片产能瓶颈论Material
2026-10-29 19:18 GMT+8

Speaking at the AI World conference in Las Vegas, Oracle Co-CEO Clay Magouyrk offered a counterintuitive assessment regarding market concerns over data center construction delays: the factor truly limiting the expansion of AI compute is not power, but chips.

He stated, 'I think everyone who spends time thinking about how much power is coming online is missing the point. You should actually be looking at how much memory is being produced and how much wafer fab capacity is available; that is a much clearer indicator because we do not need power without chips.' His logic is that even if a data center secures ample power supply, that power is useless without sufficient GPUs and supporting storage. He described this relationship using the term 'match sets': without chips, no one will lay fiber optics, purchase transceivers, or perform mechanical, electrical, and plumbing installations.

This argument directly challenges the most common narrative in current AI infrastructure discussions. The market is generally concerned about whether the grid can keep up with data center demand, whereas Magouyrk shifts the constraint from the energy side to the semiconductor supply side. He cited an example where someone emailed him claiming they had a farm in Iowa capable of hosting up to 300 gigawatts (GW) of facilities; adding that figure would bring next year's total to 340 GW—but 'that does not mean I could find any computers to put into it.'

Regarding the construction delays themselves, Magouyrk used a film production metaphor: 'Sometimes it feels like you wake up and see a headline saying the actors showed up to the set today, but they had to wait until the lighting was right before they could record their lines again. That’s how movies are made! Building large-scale projects of this magnitude is inherently complex, and I think anyone entering these projects without assuming they will be complicated is sadly mistaken.' He acknowledged that Oracle encountered power access delays at a site in Wisconsin, faced local community opposition to a project in Michigan, and last month issued a protective force majeure notice to Stack Infrastructure, a subsidiary of Blue Owl and developer of a 1,400-acre campus in Doña Ana County, New Mexico.

However, he argued that these should not be treated as anomalous signals. He explained the actual composition of data center capital expenditure: if you remove racks, network gear, fiber optics, cabling, chillers, and uninterruptible power supplies (UPS)—because all of these are movable—the remaining civil works, site grading, and substation transformer work account for a very small percentage of the total project cost. Therefore, a delay in one project can be offset by components from another project. 'Sometimes it’s a little bit like what Bob Ross would say: you have happy little accidents. You might have something delayed here, but you can use those components to accelerate a different project elsewhere.'

Magouyrk also attempted to correct the market’s understanding of marketing terms like 'up to.' He said, 'People read press releases and seem to completely ignore a very important word—up to. That word means irrelevant. I could say I am up to 100 feet tall; I am not, but I could say up to.' This comment targeted Oracle’s announcement at the same AI World conference regarding its partnership with NVIDIA to deploy Vera Rubin AI server systems at a scale measured in multiple gigawatts, as well as capacity figures promised to investors.

Fuentes:diginomica

Investigación

Hugging Face and US Universities Launch ProtoStar to Make Public Science Data AI-Ready

Tema · ProtoStar计划Material
2026-10-29 08:00 GMT+8

Hugging Face announced the launch of Project ProtoStar on October 29, 2026, an initiative bringing together research institutions, industry, and the open science community to make publicly funded scientific datasets, models, and code more accessible and reusable, particularly for AI-driven research. The founding partners include Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University, and The George Washington University, working through the Open Forum for AI (OFAI); the Astera Institute; Hugging Face as the infrastructure partner; and the Open Research Community Accelerator (ORCA).

The initiative addresses a critical gap in the approximately $190 billion annually invested by the U.S. federal government in research and development. While agencies like the National Institutes of Health (NIH) and National Science Foundation (NSF) have strengthened public-access requirements, many scientific resources remain scattered across incompatible systems or stored in static formats lacking the metadata and documentation needed for AI systems to interpret them. ProtoStar aims to encourage systematic sharing of these artifacts in "AI-ready," machine-actionable formats. Hugging Face notes that over 17,000 public datasets across core scientific domains are already hosted on its platform, with the number of new science datasets shared in the first nine months of 2026 being three times higher than in the same period of 2025.

To demonstrate the value of this approach, ProtoStar will organize a series of open challenges. The inaugural challenge, "Climate Risk Intelligence," is hosted by Carnegie Mellon University and runs from November 2 to 13, 2026. Participants will use historical weather data from 210 U.S. cities up to 2018 to predict actual conditions from 2019 to 2026, including average annual temperature, precipitation, and cooling degree days. This setup allows models to be evaluated against observed weather rather than hypothetical future scenarios, testing whether AI can effectively combine and extend public scientific resources.

For readers, this initiative represents a structural attempt to solve the "last mile" problem in AI training data: how to integrate non-technical public research outputs into computational workflows. If successful, it could provide a template for other nations and institutions to convert vast public R&D investments into efficiently reusable digital assets.

Fuentes:huggingface.co

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

Potter cites solar photovoltaics, the Bessemer steelmaking process, and Ford’s Model T to illustrate that a technology’s impact often depends on how cheaply it can be produced. Whether fusion-driven nuclear chemistry can truly rewrite the economics of isotope manufacturing remains an area of exploration. Companies such as Marathon Fusion and SHINE Technologies in the United States, along with Italy’s SORGENTINA-RF program and New Zealand’s OpenStar, are advancing related solutions, but there are no yet-verified examples of large-scale production. Potter’s stance is conditional: he considers this direction worthy of continued attention.

He concludes that enterprise digital twins are easier to achieve than consumer scanning, and the direction is already underway. But the article ends with an open question:

"What do you most want to build a digital twin of?"

Inviting reader comments rather than offering a definitive conclusion. For AI investment research practitioners, the value of this short piece lies in clearly identifying a specific limitation: when assessing whether AI assistants can autonomously handle portfolio company tasks, the convenience of document access cannot replace the acquisition of knowledge about actual operations, which currently resides largely in people's minds and has not yet been captured by any data architecture.

On the demand side, Magouyrk believes the market similarly underestimates the breadth of AI compute needs. He noted that a year ago, using AI for software development was merely an idea, but now it is reality, and 'every year you will see significant new industries turned on [to AI].' There are over 1,000 AI customers on the Oracle platform, ranging from large model providers to trading analytics firms and space startups requiring compute power. He expects that GPU supply will not exceed demand over the next three to five years.

Regarding site scale, he distinguished between the different needs of training and inference: gigawatt-scale campuses are suitable for training workloads because they require harnessing all compute capacity in one area; however, as inference demand grows, Oracle is delivering numerous 50-megawatt (MW), 100-MW, and 200-MW sites. 'Large sites can be carved up into many small allocations, but you cannot take a bunch of small data centers and stitch them together as effectively as a large campus.'

It is worth noting that these remarks come from the co-CEO of a company whose stock price has fallen approximately 50% over the past calendar year. Its Chief Financial Officer, Hilary Maxson, simultaneously hinted that capital expenditure would peak in the current and next fiscal years, with plans to raise no more than $20 billion in additional funding through fiscal 2030. Magouyrk’s arguments serve the explicit purpose of calming investor sentiment; his judgment that chip capacity is a tighter constraint than power remains a personal assertion by corporate executives and has not yet been verified by independent industry data.

Nevertheless, this analytical framework itself is testable: if AI compute deployment speeds are indeed constrained by memory chip and wafer fabrication capacity rather than power access in the coming quarters, then the market’s pricing logic, which views power as the core bottleneck, will need correction. Conversely, if chip supply remains loose while power becomes the actual choke point, Magouyrk’s judgment will be falsified.