For prefilling, DeepSeek reports that they use 32 GPUs for this stage, meaning that each GPU holds around 8 routed experts and one shared

instead of testing one configuration, we run each pruned model across a range of expert parallelism levels, from EP1 up through EP8, covering everything from a single accelerator holding the full pruned model to spreading it across 8 accelerators

each time a token is generated, it is utilizing the compute equivalent of 4 GPUs (8 routed experts ÷ 2 experts-per-GPU), not the full 144 GPUs that the model is spread across

during decoding, each individual GPU is really only ever responsible for holding and computing a small handful of experts, around 2

According to this, the DeepSeek V3 model runs with an expert parallelism degree of 144 for decoding

realistic AI serving stack is designed to batch requests aggressively and to run at a high, sustained batch size as much as possible. This assumption matters for the rest of the AI pipeline, since it means that the batch size ranges we are aiming to mimic and the throughput and power curves we build around them should be weighted toward the higher end of what we benchmark, rather than toward the low batch size regime, which would only apply to a lightly used deployment

Prefilling is the first step, which processes the input prompt to generate the initial KV (key-value) cache values in addition to the first generated token. This phase is parallelized but compute-bound (Delavande et al. 2026), which means it is limited by pure mathematical calculation speed on the accelerator and usually leads to saturation in terms of compute and power

This process is memory-bound, which means it is limited by how fast the hardware can stream model weights and cache data from memory. By default (i.e., with no optimization), this usually leads to low utilization, with a GPU generally running at 30-60% of its total capacity in terms of power draw and memory utilization

If grids decide that they need to build new power lines, then the new law says that local utilities don’t have an automatic monopoly or a federal “right of first refusal” to build those lines. Instead, grid authorities can auction off the right to build those lines.

Europe must therefore create large, concentrated buyers by design. What is needed is for its largest companies to pool their midsized commitments into contracts large enough for investors to lend against.

But this creates the next tension. To exploit these assets, Europe needs to have control over their storage and processing. This means it needs large-scale AI data centres.

China using the licences on its open-weight models as leverage in a tariff dispute

Europe has different routes to raise growth, such as removing the high barriers in its internal market. But widely adopting AI is perhaps the most promising one today. According to ECB scenarios, fast adoption of AI would add 0.3 to 0.4 percentage points a year to total factor productivity growth — the gains that come from working more efficiently rather than adding labour or capital — which has been roughly zero since 2022.

Indonesia has about 112 million users, with WhatsApp being the leading messaging platform in the country, and in Zimbabwe WhatsApp alone accounts for roughly 44–50% of all mobile internet traffic.

(At the end of 2025, BP withdrew its involvement in one of the blue hydrogen facilities at the Teesside site. A data centre is planned for the site instead.) Net Zero Teesside, a gas-CCS power plant in the East Coast Cluster run by BP and Equinor, has been unsuccessfully challenged in court over its emissions savings. The challenge was based on the idea that potential upstream emissions could significantly exceed any emissions cuts from CCS use.

Between the CCC’s sixth and seventh carbon budget advice, its recommendations for power and industry CCS capacity dropped from 46MtCO2 to 41MtCO2

The relationship between finance, technology, and war is of course not new. Bond markets first emerged in Britain in the seventeenth century to finance war with France.

What struck me about this data is that mere opposition to AI ranked down at the bottom of the list. But when data center apologists get the microphone, that’s the issue they seize on, while ignoring entirely the top concerns that people actually have. That makes sense, because there’s no getting around the resource consumption or impact on local communities, whereas calling the opposition Luddites is easy

The case for degrowth can be made even stronger on the report’s own terms. In one of the working papers (Chancel et al., 2026, Section 7), the authors incorporate the valuation of leisure and planetary habitability into a more comprehensive welfare measure they call augmented GDP. For working time, they assume that an additional hour of leisure carries the same marginal value as an hour of work; said differently, leisure is valued at the foregone hourly wage. Because the SC reaches 60k€ with 1,000 hours of annual work, against 2,000 hours for PC and PI, the 1,000 freed hours are worth a further 60k€, lifting SC’s augmented GDP to 120k€.

If a country’s macro-ecological footprint is unsustainable, sufficiency becomes an imperative of reduction (sometimes termed “ecological sufficiency” or “eco-sufficiency”). When interpreted as such, “sufficiency” is often considered a rival to “efficiency” (or more specifically, to “eco-efficiency”).[4] If any environmental impact can be written as a level of activity times the impact per unit of activity, efficiency acts on the second term, lowering the impact intensity of each unit while sufficiency works on the first, lowering the quantity of activity itself.

Core counts are a headline fixture of device marketing, but even the cheapest phones have featured eight (slow, memory-starved) cores since 2019. Speed comes from other properties; namely appropriate cache sizing, memory latency, frequency scaling, and chip architecture. Apple's core-count restraint and focus on other aspects should have been a lesson to the entire industry long before 2024.

The gobsmacking caches of A-series parts allow Apple to keep the beast fed, leading to fewer stalls and more efficient use of CPU time. This advantage redounds to better battery life because well-fed CPUs can retire work faster, returning to lower-power states sooner. That it has taken this long for even the top end of the Android ecosystem to begin to respond is a scandal.

There are actually a lot of interesting technical details about how Apple rebuilt Siri, including expanding Private Cloud Compute to include Nvidia chips running in Google data centers, as well as a 20 billion parameter on-device mixture-of-experts model that selects the expert on a per-query basis (as opposed to on a per-token basis) so that it can run in an iPhone’s limited memory

The production of biofuels also involves fossil fuels in multiple ways. It starts with nitrogen fertilizers made from fossil gas. Refinery processes like hydrogenation use Hydrogen, and that is, these days, by and large also made from fossil gas. (Hydrogenated Vegetable Oil, or HVO, a common type of biofuel, even has this in its name.)

Here, we propose a solvent-mediated molecular engineering strategy to fabricate high-performance bamboo molecular plastics (BM-plastics) with tunable H-bond networks. By employing a hydrated ZnCl₂/formic acid deep eutectic solvent (DES), we disassemble the native H-bond matrix of bamboo cellulose into a homogeneous molecular system. Subsequent ethanol stimulation triggers the rearrangement of cellulose chains, fostering dense, ordered H-bond interactions between hydroxyl and formate ester groups (Fig. 1b). This dual-step process not only overcomes the trade-off between mechanical strength and processability (Fig. 1c) but also enables scalable manufacturing under ambient conditions (Fig. 1d). In contrast to traditional processing methods that often require toxic or costly reagents and are associated with high energy consumption, our solvent-mediated molecular engineering approach delivers substantial operational advantages while significantly improving the performance characteristics of the resulting BM-plastic

Firms that can show this level of granularity will be in a much stronger position to reduce emissions exposure under CBAM and stay competitive in the EU market

Each scenario was solved with Gurobi v.12.0.1 on a high-performance cluster node with two AMD EPYC 7543 CPUs and 16 DDR4-2933 64GB ECC RAM units. With 8 threads available, a single scenario required between 16 and 19 h and up to 130 GB of memory

In both pathways, biogenic carbon and hydrogen present in the syngas are used for methanol synthesis

Since the commercialization of methanol-to-olefin (MtO) technology in 2010, China has been using methanol synthesized from coal for a large share of its olefin production.

it can power gas turbines with minimal retrofitting2424.Cui, J. ∙ Yu, R. ∙ Wang, H. ...A Comparative Study of Methanol and Methane Combustion in a Gas Turbine CombustorEnergies. 2025; 18:1765Google Scholar (conversion costs are estimated at $3 million per turbine,2323.Bertau, M., Offermanns, H., Plass, L., Schmidt, F., and Wernicke, H.J., eds. (2014). Methanol: The Basic Chemical and Energy Feedstock of the Future: Asinger’s Vision Today, First Edition (Springer). https://doi.org/10.1007/978-3-642-39709-7.Google Scholar less than 1% of the investment cost of a 500 MW turbine);

Greystoke, a company planning to build another datacentre in north Lincolnshire, one of the largest in the UK, also appears to have misstated the emissions of its project in the same way. Taken together, the three developments will account for more than 1% of the UK’s carbon budget in 2033. This is the equivalent to the emissions of a mid-sized city such as Bristol.

His medium-term market strategy combines product placement with policy advocacy. In the early phase, Kaypee Motors is prepared to absorb a loss on each vehicle sold, selling below production cost to build market visibility and begin the trust formation that a new local brand requires in an environment where most consumers have never seen a Tanzanian-manufactured vehicle. "You may spend $10,000 to produce a car and sell it at $9,000," he said. "That means you are taking a loss. But sometimes that loss is part of market penetration. You want people to see the cars, see the vehicles moving, working, carrying goods, and solving real problems." Over the medium term, he believes that policy intervention, specifically the use of import tariffs calibrated to reduce the volume of used vehicle imports and create protected market space for local producers, is the structural complement to entrepreneurial effort that no amount of individual innovation can substitute. "When local manufacturing is more ready, I believe we can lobby the government to create policies that favour local production," he said.

The company he has built toward that conviction does not yet resemble what most people imagine when they hear the phrase electric vehicle manufacturer. The factory he has built toward that conviction does not yet resemble what most people imagine when they hear the phrase electric vehicle manufacturer.

“That’s what we’ve seen happening in Virginia, where every data center job creates six jobs in the local economy,” Vaquero said. Such figures are contested by economists, who argue they are based on modeling assumptions that overstate real-world impacts and conflates construction jobs with permanent ones.

However, this notion is not confined to grid-constrained markets. France, which had a surplus grid capacity originally planned for transport and heating, is finding that AI clusters have been faster to claim, displacing the very sectors the headroom was intended to serve. 150

Indeed, it is technically possible to dynamically reallocate idle reserved capacity to other tenants, as modern AI cluster management systems can automatically assign other workloads to GPUs that are temporarily idle, and production deployments have demonstrated that dynamic sharing can increase utilisation from less than 30% to more than 70%. 137 However, the stronger the access guarantees owed to capacity holders, the less freedom the scheduler has to fill their idle periods with other workloads; we call this ‘the self-hosting trap’. Public law entities expecting guaranteed on-demand access cannot accept their reserved capacity being occupied by other jobs at the moment they submit their own workloads

The self-hosting arrangements embedded in the AIGF governance structure also introduce a governance constraint on scheduling flexibility. Under the EuroHPC Regulation, the Union’s financial contribution, capped at 17% of the capital expenditure, entitles it to proportional access, with priority for public law entities, industrial users on EU-funded projects and SMEs

A multi-client AIGF producing sub-second power swings of this magnitude is therefore actively increasing the system’s balancing costs, not merely underutilising its connection

The hardware provided for those peaks continues drawing power through the quiet periods. Independent grid data corroborates the scale of the gap. Ember’s analysis of European TSO data revealed that data centres use, on average, only 44% of their contracted grid capacity, using 34% of it in Ireland and 30% in Norway

The multi-client scenario assumes a fragmented user base of enterprises, SMEs, public institutions and research organisations with heterogeneous, largely inference-driven workloads. 123 The paper concludes that given Europe’s provider landscape and user base, the anchor customer scenario is improbable and the multi-client scenario is far more realistic. This chapter takes that conclusion as its starting point. The following subsections show how a multi-client AIGF structure sharpens three bottlenecks—the utilisation gap and its energy costs, load variability and system-level impacts, and scheduling complexity—and contrast these outcomes with the anchor customer model to illustrate what the same GPU hardware would look like under a different use case.

At the chip level, utilisation refers to the share of a GPU’s theoretical compute capacity that is actively performing useful work at any given moment. In tightly optimised benchmark or hyperscale training runs, cluster-level GPU utilisation can reach approximately 90% for extended periods once data pipelines and storage bottlenecks are carefully engineered. However, this figure represents the best-case scenario. Studies of commercial deployments where workloads are more varied and scheduling policies face limitations find that utilisation is considerably lower at 30–70% on average. At the cluster level, utilisation takes on a different meaning, as it describes how much of the installed compute capacity is actively drawing load over time, as opposed to sitting idle or in standby

Within each AI server, GPUs are linked by ultra-high-bandwidth, short-reach interconnects such as NVLink. Instead of sending data over the standard connection that normally links a GPU to the rest of the server, NVLink creates a direct ‘fast lane’ between GPUs so that they can exchange model data without going through the CPU, as it is the case for the standard connection. 61 This high-speed fabric effectively pools the memory and compute of multiple GPUs inside a rack but also raises local power density, adding extra heat that needs to be removed at rack level

At the same time, improvements in power semiconductors, such as gallium-nitride–based converters and very efficient DC-to-DC direct current transformer modules, make each remaining conversion stage waste less energy and heat

Here’s the twist: in the last 10 years, battery costs dropped from $440 to $70/kWh. 45% in 2025 alone!

The project is huge but the tl;dr is that we're describing "AI safety" as technology that is safely created, safely deployed, and safely used and maintained. We're setting a high bar, and it kicks in way before concerns about existential risk become material. Is your technology made in ways that exploit human labour and disregard planetary boundaries? If so, there's a strong argument to be made that it's not safe and should be subject to appropriate mitigations.

One of the things coming through at this stage of the project is how difficult it is to maintain a vibrant and plural AI ecosystem, or set of ecosystems, and enable different approaches to development. The dominance of the big labs and preoccupation with AGI means that most “responsible” efforts get distracted with tidying up after obvious mistakes. As such, there’s relatively little time or funding left to invest in alternative approaches such as pursuing computing within planetary boundaries or building capabilities around small AI.

but it’s really a choice about direction

New data centres connecting to the electricity network will continue to be required to provide generation and/or storage capacity, either onsite or local in proximity to match the requested data centre maximum import demand capacity

Data centres will be required to meet at least 80% of their annual demand with additional renewable electricity projects generating in the Republic of Ireland. To ensure additionality, renewable generation which has previously been, or is currently, contracted through existing support mechanisms (e.g. REFIT, RESS, ORESS) cannot be used to meet this requirement. Data centre developers will be allowed a 6-year glide path for renewables to be developed and start generating. This acknowledges the timelines associated with developing renewable energy projects. Data centre developers will be required to present a credible plan to System Operators (EirGrid and ESB Networks) for the development of associated renewable projects.

The ending of the pause took an entire new paradigm to kick off. Railroads became a new driver of labor demand that took the slack that the industrial revolution created.

Regulatory Coordination: Emphasizes the need for coordination between the PUC, ERCOT, TCEQ, and EPA to provide enforcement discretion for air emissions limits when large loads use backup diesel generators during grid emergencies

We sustain our development velocity on Ruff, uv, and ty through extensive CI/CD workflows that run on GitHub Actions. Without these workflows we would struggle to review, test, and release our tools at the pace and to the degree of confidence that we demand. Our CI/CD workflows are also a critical part of our security posture, in that they allow us to keep critical development and release processes away from local developer machines and inside of controlled, observable environments.

Unfortunately, there's a flipside to this: GitHub Actions has poor security defaults,

Pakistan now sometimes produces more electricity from rooftop solar than from its main power grid

Posting on LinkedIn, Sam Hollister, LCP Delta principal and head of UK market strategy, writes that a spike in gas prices similar to current levels would only increase household bills by 8%, if the 2030 targets are met. In contrast, bills would rise by 45%, if no CfD-backed renewables were on the system.

If the world can survive this energy transition and the fossil fuel wars soon to be followed by water wars, the U.S. will have voluntarily relinquished its position as a world leader and handed the keys to China

Can you blame them?  Maybe someone would be more willing if there hadn’t been an LNG tanker taken out with an unmanned drone recently. Ships full of hydrocarbons are pretty risky when it’s raining explosives.

Last June, Allen and Bannon lobbied against the industry’s attempt to pass a 10-year federal moratorium on state regulation of AI, putting in calls to Republican Senators like Marsha Blackburn—who pulled her support from the bill at the 11th hour, causing it to fall apart. (Trump issued an Executive Order in December to challenge state AI laws.)

Trump is using tech companies to attack his geopolitical rivals, ordering Microsoft to shut down both the International Criminal Court and a Brazilian high court in retaliation for their pursuit of the criminal dictators Benjamin Netanyahu and Jair Bolsonaro. This means that Trump has violated the quid pro quo deal for keeping anticircumvention law on your statute books, and he has made the case for killing anticircumvention as quickly as possible in order to escape American tech platforms before they are weaponized against you

The real annoying thing about Opus 4.6/Codex 5.3 is that it’s impossible to publicly say “Opus 4.5 (and the models that came after it) are an order of magnitude better than coding LLMs released just months before it” without sounding like an AI hype booster clickbaiting, but it’s the counterintuitive truth to my personal frustration

Here’s the kicker: none of this costs anything at runtime. Mojo's compile-time metaprogramming means the abstractions exist only in the source code. The GPU assembly the compiler produces is identical to what you'd get from a hand-written, monolithic kernel

This paper examines whether Biden-era green investments affect public opinion in ways that could sustain and expand climate reforms. This process is known as policy feedback, where public policy creates supportive constituencies (6, 7). Our focus on public opinion, rather than electoral outcomes, allows us to assess whether the political logic motivating these investments operates at its most basic level, prior to voting or legislative conflict

For policies to affect politics, voters need to be able to trace them back to the responsible political actors, which is challenging in a complex information environment

Long-distance communication latency is dominated by the length of the network and the propagation speed of fiber. Modern fiber optic can propagate signal at two-thirds the speed of light, for a latency of 5us/km.17 This would make our 4,800km network have a latency of 24ms.18

Comrie wrote a 1944 Mathematical Gazette article entitled “Careers for Girls,” in which he declared that female computers were useful “in the years before they (or many of them) graduate to married life and become experts with the housekeeping accounts!”

OpenClaw, like many other open-source tools, allows users to connect to different AI models via an application programming interface, or API. Within days of OpenClaw’s release, the team revealed that Kimi’s K2.5 had surpassed Claude Opus and became the most used AI model—by token count, meaning it was handling more total text processed across user prompts and model responses.

Meanwhile, under China's 'Eastern Data, Western Computing' initiative in the early 2020s, numerous startups constructed large AI and cloud data centers across western regions of China, where electricity costs are lower, with the goal of serving demand from economically stronger eastern provinces. While the strategy reduced power expenses, it turned out that longer distances increased latency and made these facilities less attractive for many latency-sensitive applications, which limited actual usage

Frontier AI labs are running out of CPUs for their RL Training needs and are scrambling for CPU allocation by competing directly with the cloud providers for commodity x86 CPU servers. Intel, facing the unexpected depletion of their CPU inventory, is looking to raise prices across their Xeon line while they ramp additional tools to shore up CPU production

In addition, our Li-ion BBUs benefit from our distributed UPS architecture that offers significant availability and TCO benefits compared to traditional monolithic UPS systems. The distributed UPS architecture improves machine availability by: 1) reducing the failure-domain blast radius to a single rack, and 2) locating the batteries in the rack to eliminate intermediate points of failure between the UPS and machines. This architecture also provides TCO benefits by scaling the UPS with the deployment, i.e., reducing day-1 UPS cost. Additionally, locating the batteries in the rack on the same DC bus as the machines eliminates intermediate AC/DC power conversion steps that cause efficiency losses. In 2016 we shared the 48V rack power system spec with the Open Compute Project, including specs for the Li-ion BBUs.

At Google, we rely on a 48Vdc rack power system with integrated battery backup units (BBUs), and in 2015, we became one of the first hyperscale data center providers to deploy Lithium-ion BBUs. These Li-ion batteries had twice the life, twice the power and half the volume of previous-generation lead-acid batteries. Switching from lead-acid batteries to Li-ion means we deploy only one-quarter the number of batteries, greatly reducing the battery waste generated by our data centers. 

The lithium-ion batteries that China dominates are becoming increasingly prevalent. In February Google said that it had installed more than 100 million cells across its data centers and had started to replace diesel generators with batteries. Microsoft said it aimed for its data centers to eliminate diesel fuel for backup by 2030 to meet environmental goals.

Airbnb picks Alibaba’s Qwen over ChatGPT in a win for Chinese open-source AIAirbnb ‘relies heavily’ on Alibaba’s Qwen models to power its AI customer service agent, CEO Brian Chesky says

Low-cost Chinese AI models forge ahead, even in the US, raising the risks of a US AI bubble Nvidia’s latest earnings report reassured some. But Chinese AI models are fast gaining a following around the world, underlining concerns over an ‘AI bubble’ centered on high-investment, high-cost US models.

In Belgium, we’ll soon install the first ever battery-based system for replacing generators at a hyperscale data center. In the event of a power disruption, the system will help keep our users’ searches, e-mails, and videos on the move—without the pollution associated with burning diesel. But even more important is what will happen when Google doesn’t need emergency power. Whereas diesel generators sit idle most of the year, batteries are multi-talented team players: when we’re not using them, they’ll be available as an asset that strengthens the broader electric grid.

A key challenge when deploying off-grid generation is managing redundancy. For example, the 1.4GW Vantage DC campus in Shackelford County, TX will deploy 2.3GW of VoltaGrid systems. These systems being small facilitates redundancy – but if you were to deploy onsite power with large heavy duty turbines, redundancy might be to simply have two power plants, if not more.

Maintenance has historically also been notably higher than other solutions. Individual fuel cell stacks last roughly 5-6 years, then must be replaced and refurbished. This per-cell replacement makes up ~65% of service costs, although specific numbers are kept close to vest. Bloom discloses little about its materials beyond the use of ceramics in the cell core, but claim that their fuel cells have no critical mineral dependence on China or other contested regions

High-speed engines can be much faster to deploy. For example, at the initial Colossus 1 deployment, xAI leveraged 34 VoltaGrid truck-mounted systems, incorporating high-speed engines from Jembacher. High-speed engines, in particular, are popular with energy procurement vendors (described later). Their wide availability and small unit size offer faster time-to-power. We show below a VoltaGrid 50MW deployment in San Antonio, with twenty Jembacher J620 (rated 3.36kW per unit

AI infrastructure developers cannot wait five years. In many cases, they cannot wait six months, because waiting six months costs billions of dollars of lost opportunities.

However, AI infrastructure cannot wait for the grid’s multiyear transmission upgrades. An AI cloud can generate revenue of $10-12 billion dollars per gigawatt, annually

The system cost is higher than the solar levelised cost because it also includes the costs of the batteries and backup power, as well as accounting for curtailment. Utility-scale solar PV and batteries are sized to minimise the costs for meeting a constant demand based on hourly weather data from 2011. The 1° by 1° pixels are chosen based on where the population is densest, representing where 99.9% of the population lives. The model is based on model.energy but without the hydrogen storage. The most important economic data is based on the Danish Energy Agency Technology Database for 2030 in 2020 euros: solar PV installed cost 384 €/kWp, lithium-ion battery installed cost 157 €/kWh. By 2050 these reduce to 293 €/kWp and 83 €/kWh respectively. The solar costs include 50 €/kW for transformer and grid connection, but no further grid costs are included. Additional assumptions: cost of capital 5%, cost of backup generator with 50% efficiency 1000 €/kWel, cost of backup fuel (based roughly on fossil gas) 30 €/MWhth. There is more information below including instructions on how to adjust the results if you don't like the inputs

The graph shows the results for 90%, 95% and 99% solar-battery supply

So what McKinsey fails to mention is that LNG is competing with coal and renewables and while $7-$10 is in the range of pricing where U.S. LNG exporters lose money, it is still too high to compete with coal and renewables in the major Asian markets

Demand will outpace growth. And yet the LNG export terminals will need a lot more gas to fulfill the existing contracts in place over the next five years. Where will this come from? Most of it is expected to be from the Permian as associated gas from oil production or from the Haynesville from gas wells

Elsewhere, Israel’s Nostromo Energy became the latest vendor to bring a data center “ice battery” to the market. Ice batteries work by cooling ice stored in cells at times where demand for power is low. The batteries use the cold temperature of the ice to chill the cooling liquid needed in HVAC systems, eliminating the need for mechanical chillers, which are big consumers of electricity. Nostromo believes it can make up to 40 percent of data center power consumption flexible, which could be a big boon for operators looking to trim costs.

Open standards are emerging in liquid cooling, and many vendors are following the blueprint laid out by Google with its Project Deschutes, the in-house CDU design developed by the cloud and search giant. Revealed earlier this year, Google has made the specs of Project Deschutes available to the Open Compute Project (OCP) community.At the OCP Europe Summit, held in Dublin in May, Google revealed the extent of its own liquid cooling deployments, demonstrating how the technology has taken hold among the hyperscalers. Around half of Google’s global data center footprint has liquid cooling enabled and/or deployed, meaning the firm has around 1GW of liquid cooling capacity deployed across 2,000 pods equipped with its Tensor Processing Unit AI chips. It achieved uptime of 99.999 percent.

Second, a volume mechanism: even the cost-optimal procurement strategy for 24/7 hourly matching might include some amount of excess clean energy. If total CFE generation of assets procured by 24/7 participants exceeds demand in a given hour, the “excess CFE”is not counted towards a CFE target; however, it is clean electricity that can potentially be stored (using batteries or LDES), or sold to the regional grid. As for the latter, excess CFE might displace emitting grid generators or reduce the need to import electricity from neighbouring areas.

First, a profile mechanism: participants doing 24/7 hourly matching procure clean energy resources that better match their demand patterns. As certain consumers align their demand with CFE supply on hourly basis, the need for dispatchable generation that can firm intermittent renewable supply is lower in the rest of the system. This mechanism can reduce the utilisation of fossil-based generators, such as OCGT power plants, which typically ramp up when wind and solar resources are scarce.

Bores helped write, pass and just secured Hochul’s signature on the RAISE Act, a sweeping piece of legislation aimed at regulating the industry. But it passed only after the pro-business governor struck the entirety of the original text and replaced it with a new draft that reportedly copied almost verbatim a California bill that is much more friendly to AI companies. After an intense back and forth, she ultimately signed a bill that was somewhere in between the original draft and her edit

The Texas Permian is the story of U.S. shale and it has peaked. In October of 2019 Texas oil production topped 165 million barrels. In October 2024 it topped 180 million barrels. In August 2025 it was back to levels first hit in 2019 and the initial numbers for September 2025 were 144 million barrels.

I’ve yet to see one acknowledgement in the major financial press about how badly the Bakken performed compared to how it was sold. Why should this happen?  Because the average American still thinks there is endless oil and gas in the U.S. shale. They are going to pay the price for the industry fraud in multiple ways. First, with higher electricity and heating prices as natural gas supplies become scarce. Secondly, they don’t realize how the current administration’s attempts to block renewable energy and electric vehicles in America is based on the lie that there is plenty of cheap oil and gas to use instead.

In 2020 I wrote a piece on how the Bakken oil production was peaking and what this would mean for the asset retirement obligations for the Bakken. In that piece I pointed out another aspect of the Bakken that was getting little attention. More than a decade after the Bakken boom started, North Dakota was flaring 23 percent of the methane gas produced via fracking — making a mockery of the state’s flaring regulations. Harold Hamm, Bloomberg and many others talked a lot about technological innovation while the reality in the Bakken of North Dakota was that the industry was wasting almost a quarter of the natural gas being produced. For comparison, the target the European Union is currently using for its methane is 0.2%.  Also for comparison, Iraq is flaring around 25% now.

No direct access to user prompts or model outputs was available for this study. Instead, OpenRouter performs internal categorization on a random sample comprising approximately 0.25% of all prompts and responses through a non-proprietary module GoogleTagClassifier.

Use locational pricing to manage power grid congestion (in countries where this is currently politically difficult: use tools for locational investment signals, which can provide much of the benefit of locational prices [JRC, 2025]; vary network charges spatially and temporally to get operations right).

Build new dispatchable backup power plants (turbines, engines, fuel cells) to run on carbonaceous fuels (easily switchable from methane to methanol, fossil gas being displaced over time by e-biofuels); don't add cost by insisting they be hydrogen-ready. (Germany needs 70-90 GW of dispatchable capacity, depending on which study you look at.)

Changing plastics end-of-life policy to managed landfilling would avoid the need to provide green primary high value chemicals (HVC).

The extra cost of going down the e-biofuel route instead of compensating fossil fuels with bio-CDR would be manageable. To replace 400 TWh/a gas demand and 1200 TWh/a oil demand in Europe with e-biomethanol rather than bio-CDR: (400-200) EUR/tCO2 * 80 MtCO2/a + (240-200) EUR/tCO2 * 300 MtCO2/a = 28 bnEUR/a. This is around 0.1% of future European GDP; it is in a similar range to current European subsidies for first generation biofuels.

If advanced e-biofuels are the more scaleable option, then they have to be produced everywhere, including in Europe, because supply is scarce. It would have several other advantages: replacing farmer's income streams as traditional 1st generation biofuels are phased out; and anchoring value creation in rural areas for decades to come.

In many countries straw is burned without any energetic use at all, causing widespread health harms from particulate matter, so alternate uses may bring advantages.

An example: for a hydrogen peaking power plant in Germany running 200 hours a year, the capped network capacity charge for withdrawing from the planned Kernnetz pipeline network of 25 EUR/(kWh/h-peak)/a [BNetzA, 2025] works out at 25 EUR/kW/a / 200 h/a = 125 EUR/MWh ~ 4 EUR/kg. Add this to a production cost near Germany of 120 EUR/MWh and a storage charge of 120 EUR/MWh [EWI, 2024], and you are quickly at 360 EUR/MWh for the fuel alone.

As we are on the precipice of a very large wave of lending, I also have to ask myself, is capitalism itself ready for it? More thoughts behind a paywall

To unlock this flexibility, Germany must expand storage options and ensure users can respond to power availability. A key step came with the obligation for electricity suppliers to offer at least one dynamic tariff from 2025, but additional incentives must follow. As Germany paves the way towards covering 80 percent of its electricity demand with renewables by 2030, measures and proposals for 2026 to unlock flexibility include

During most of the year, our grid has a lot of excess capacity. The peak demand in the United States is generally less than 800 gigawatts. The average demand on our grid most of the year is about 400 to 450 gigawatts. So most of the time you’ve got 200 to 300 gigawatts of spare capacity in the grid. That’s more than enough to meet all the data center growth and EV growth and new manufacturing and all the things we’re doing in our country.

The agent discovers tools by exploring the filesystem: listing the ./servers/ directory to find available servers (like google-drive and salesforce), then reading the specific tool files it needs (like getDocument.ts and updateRecord.ts) to understand each tool's interface. This lets the agent load only the definitions it needs for the current task. This reduces the token usage from 150,000 tokens to 2,000 tokens—a time and cost saving of 98.7%.Cloudflare published similar findings, referring to code execution with MCP as “Code Mode." The core insight is the same: LLMs are adept at writing code and developers should take advantage of this strength to build agents that interact with MCP servers more efficiently.

Just as Big Oil has repeatedly failed to deliver on pledges to begin decarbonising, so too the promises of plastics companies have been hollow. In the early 1990s, for example, Coca-Cola announced its intention to make its bottles from 25 per cent recycled plastic, only to quietly ditch the target four years later, after political and consumer pressure had eased.

PJM’s initial proposal to deal with the data center swell would have created a category for new large sources of demand on the system to interconnect without the backing of capacity; in return, they’d agree to have their power supply curtailed when demand got too high. The proposal provoked outrage from just about everyone involved in PJM, including data center developers and analysts who were open to flexibility in general, who said that the grid operator was overstepping its responsibilities.

With revenue equivalent to 17.3 percent of GDP in 2022, the lowest in the OECD, the state lacks the capacity to finance social programs and strategic investments. Public investment in science, technology and innovation (STI) has particularly suffered, reaching just 0.27 percent of GDP in 2022, compared to an OECD average of 2.73 percent. Mexico stands far below economies in South Korea (5.21 percent), Germany (3.13 percent), or Brazil (1.15 percent in 2020

Far from representing an obstacle to Sheinbaum’s climate objectives, it is precisely through SOEs that the President is pursuing her vision for the green transition. Her administration has reinforced AMLO’s program to revive Mexico’s SOEs—Pemex and CFE—with new decarbonization and industrial policy ambitions. On some policy issues, the difference is one of means rather than ends. For example, in order to increase the share of low-carbon electricity by 10 to 20 percent—a commonly agreed policy aim—Sheinbaum’s government has tasked SOEs with orchestrating investment, and turned to state regulatory capabilities to ensure private sector investment. In other cases, the ends themselves have become more ambitious: in the absence of private investment in the EV market, the government has initiated a new business venture, Olinia, to manufacture small urban EVs and launched a lithium battery industry

Critics of the active state in Mexico allege that market intervention generates underinvestment. Furthermore, they claim that Mexican state-owned enterprises (SOEs) are not adequate to competently navigate new waves of technology. And, in the wider North American context, SOEs are thought to put Mexico’s industry in jeopardy by triggering mechanisms of investment protection.  The suspicion of SOEs—whether in general, or specific to the green transition—is unfounded. Most private investment in renewable energy across the globe has been facilitated by long-term supply contracts from SOEs or central government agencies. Out of all solar and wind energy projects, two-thirds are held by Chinese SOEs (incidentally, China has recently confirmed an impressive target of 3,500 Gigawatts of solar and wind by 2035). And new forms of intervention in the energy sector are welcome by investors searching for low-risk investments with long-term return profiles

Unlike traditional approaches that require training separate specialized models for each task – one for crop mapping, another for deforestation, and another for land use classification, among others – OlmoEarth excels across a wide and diverse range of Earth observation applications

DR Tulu has access to three search tools: google_search, which returns top web snippetsweb_browse, which extracts full-page text from URLspaper_search, which retrieves relevant paragraphs from open-access research papers

Top left shows electric 2 wheeler sales in China. And then the rest of the world is below that. I would expect we are going to see a lot more demand destruction from 2-wheelers in the rest of the world in the next 5 years. However, it is pretty clear that one of the technologies that drove early oil demand destruction (that first 1 million bpd by 2018), is many times the size it was in 2018 and increasing everywhere but China.

For instance, a recent analysis by Epoch AI of the total training cost of AI models estimated that energy was a marginal part of total cost of AI training and experimentation (less than 6% in the case of all 4 frontier AI models analyzed), and a recent analysis by Dwarkesh Patel and Romeo Dean estimated that power generation represents roughly 7% of a datacenter’s cost.

Chinese officials predict widespread V2G adoption by 2030 would unlock 1 billion kilowatts of capacity from an expected fleet of 100 million EVs. The move could help diversify energy sources beyond coal in a country that’s home to 40 million electric cars.

Tesla’s China-made deliveries rose by 2.8% from a year ago in September after two months of decline, boosted by the China-exclusive six-seater Model Y L launched in August.

While closed-circuit cooling systems (i.e. where all of the water is recycled and none of it evaporates)[33] are technically feasible, they are more costly and therefore less common.

While new reactors are continuing to be built globally, this takes, on average, 6 to 8 years [23];

At the level of a hyperscale data center cluster, this can translate into requirements of up to 5 and even 10 GW of power, up from 5 MW - a 2,000 fold increase in the span of a decade [4, 11].

They are very geographically concentrated - only 32 countries have data centers, and nearly half of them are in the United States. The state of Virginia has the highest density of data centers globally - it is home to almost 35% of all hyperscale data centers worldwide.

Design is the process of trying to address a problem with a new feature. Standardisation is the process of documenting consensus.

With DSQL you don’t need to pick hardware, manage clusters, design for failover, think about patching engines or systems, monitor CPU and memory, or any of the many other tasks that come with good cluster management. When you get started with any database, you can basically ignore these things: buying excess CPU and memory at small scale is cheap, and monitoring is easy when you aren’t also trying to drive efficiency

What should be discarded is the idea, prevalent in the mainstream media worldwide and even at the UN, that charging for the social cost of carbon and its “offsetting” can replace public planning for socio-technological transition.

the scale required for carbon market offset projects—through environmental conservation, forest regeneration, or reforestation—to ensure net-zero emissions is absolutely unfeasible. Offsetting cannot function as an alternative to decarbonizing energy, production, and infrastructure systems simply because there is not enough land to replant forests.

The strategy for transitioning to renewable energy should not, therefore, begin with raising carbon prices and waiting for the market to offer an alternative solution in the short term. Instead, it must involve government planning for the supply of new green technologies and substitute infrastructure based on sustainable energy. Only after these technologies and infrastructure are accessible can raising carbon prices induce rapid substitution. Until then, consumers and user companies will be hostage to dirty options

In practice, energy’s inelastic demand means that even if prices rise, users of dirty technologies and fossil fuels will continue to pay, even if they become poorer, because they cannot find feasible substitutes

The global carbon market approved in Baku will not overcome this problem. International competition to reduce costs to attract private investment is a systemic feature of the world economy. It creates a coordination problem, making it impossible to establish a single international price for carbon. International divergence in carbon prices undermines an effective system for reducing global emissions

Uruguay’s carbon tax came closest to the IPCC figure, pricing the emission of one metric ton of greenhouse gases at $167.17. At the other end of the spectrum, however, Indonesia’s carbon market charged only $0.61 for the same ton.

In a recent paper, researchers from Harvard and the University of Pisa reported that “U.S. data centers produced 105 million tons CO2 equivalent gasses in the past year with a carbon intensity 48 percent higher than the national average.”

Ironically, the policies meant to unlock innovation have ended up reinforcing Li-ion’s dominance. Italy offered 15-year fixed contracts, exactly the kind of revenue certainty newer techs need. But only lithium-ion could hit the bankability and pricing targets. In the UK, Ofgem’s eligibility screen knocked out most alternatives before they could even bid.

For Linux distributions to be able to do all of these things with the T2, Apple would need to either open up access to it, or someone in the Linux community would need to reverse-engineer support for it. Neither of those things has happened. And given the age and relative obscurity of the T2 hardware, it doesn’t seem enormously likely at this point.

However, xAI has been clear that this control technology has not been installed. An xAI consultant told The Daily Journal that the turbines would have to be retrofitted to install them. No one has claimed credit for the mailer sent by “Facts Over Fear.”

Since the turn of the millennium, the global LNG market has absorbed every supply wave fairly quickly, taking two to three years. China swallowed a good chunk of the 2009-2011 wave, when supply jumped by about 40% after the completion of several projects in Qatar. Europe absorbed the 2016-2019 wave, which came after a huge buildup in US export capacity increased global production by 45%.Now, a third wave, which is likely to extend from 2026 to 2030, looms. It’s by far the largest, potentially adding a further 60% to the world’s supply. Thus, demand is key

Gas turbines draw in air that is compressed as it passes through a series of rotors and blades, and the compression heats it to about 500C. That hot air is used to burn natural gas, lifting the combusted gas output’s temperature further to around 1,500C

The rapid rise of AI in the US has surprised utilities that had planned for relatively flat growth in demand. BloombergNEF forecasts that data-center power consumption will more than triple by 2035. Electricity needs from these facilities will account for almost half of the growth in total power demand between now and 2030, according to the IEA.

Coal-dependent Asian nations could cut power sector emissions by at least a third by the middle of next decade by expanding the use of liquefied natural gas (LNG), according to a June study commissioned by the Asia Natural Gas & Energy Association, an industry group that counts Exxon Mobil Corp. and Chevron Corp. among its members

Kasirzadeh’s account of accumulative risk still relies on threat actors such as cyberattackers to a large extent, whereas our concern is simply about the current path of capitalism. And we think that such risks are unlikely to be existential, but are still extremely serious

Access controls prevent people working with sensitive data and systems from accessing confidential information and tools that are not required for their jobs. We can design similar protections for AI systems in consequential settings

But this process took over two decades instead of a few hours in the case of AlphaZero because safety considerations put a limit on the extent to which each iteration of this loop could be scaled up compared to the previous one

What will it take for AI to push the boundaries of such knowledge? It will likely require interactions with, or even experiments on, people or organizations, ranging from drug testing to economic policy. Here, there are hard limits to the speed of knowledge acquisition because of the social costs of experimentation. Societies probably will not (and should not) allow the rapid scaling of experiments for AI development.

The Ladder of Generality in Computing. For some tasks, higher ladder rungs require less programmer effort to get a computer to perform a new task, and more tasks can be performed with a given amount of programmer (or user) effort.

But this comparison does not account for differences in the intensity of adoption (the number of hours of use) or the high cost of buying a PC compared to accessing generative AI.

the impact of AI is materialized not when methods and capabilities improve, but when those improvements are translated into applications and are diffused through productive sectors of the economy.

In reality, Arizona already has large numbers of data centers throughout the state. They collectively use about 1/50th as much water as just the golf courses in Arizona, but generate more tax revenue than the golf industry. Building data centers in the desert has been normal for a while and doesn’t harm water access. Here’s a comparison2 of water usage for data centers and golf courses in the county in Arizona with the most data water draw

We are also evolving our software stack, including Twine and MAST, to support long-distance training across a geographically distributed set of data centers

This variation by resource theme means that field-recovery upside is unevenly distributed between countries. Iran, Venezuela, Iraq, and Russia stand out with the largest top-quartile reserves upside of any countries.

Quite how this potential will be realised is another matter. The existence of recoverable reserves in the ground does not mean that these will be economically attractive or that the industry will make the investment required.

Xu et al. (2023) bring this discussion from the realm of arguments into quantitative analysis, concluding that “emissions matching strategies have zero or near-zero long-run impact on system-level CO2<math><msub is="true"><mrow is="true"></mrow><mrow is="true"><mn is="true">2</mn></mrow></msub></math> emissions” [40]. The reason for this, citing the authors, is that voluntary carbon-free energy procurement under this strategy almost exclusively displaces other carbon-free energy rather than fossil fuels.

One of the simplifications of mathematical models implementing 24/7 CFE procurement is how the excess carbon-free electricity is handled. The “excess CFE”can be defined as the total amount of carbon-free electricity supply exceeding the demand of the participating consumers in a given period of time. Here we assume that the excess can either be curtailed or sold to the regional electricity market at wholesale market prices, until the excess constraint is reached. The constraint restricts the volume of excess electricity sold to the regional grid, ensuring it does not surpass 20% of the annual demand from participating consumers.

Fig. 7(b) compares reductions in annual zone emissions in percent points achieved through 100% annual renewable matching and 100% 24/7 CFE matching; no procurement policy is assumed in the counterfactual scenario.

The figure demonstrates that, even though the absolute impact of procurement policies diminishes over time, the relative impact of 24/7 CFE on system-level emissions increases over time when compared to 100% RES

First, a profile mechanism

Second, a volume mechanism

One of the major difficulties in getting things done in a large organization is the fact that it is practically impossible to pinpoint responsibility. A Bureau or an Office can't be made responsible because, as I said before, they are inanimate and therefore are incapable of perception or of feeling. Limited tours of duty of two or three years in positions of responsibility accentuate this situation.

Instead of requiring each power plant to complete lengthy studies of needed system-wide infrastructure investments before connecting to the grid, the “connect and manage” approach in Texas gets power plants online quicker than a “studies first” approach

Azure to access OpenAI services because it allows specifying regions to use, which is important when serving EU customers, for example. Using OpenAI’s services directly would not allow switching of regions.

The team had the idea to create an “approved answers” knowledge base for the chatbot, and faced the challenge of creating the questions for this. The team made the chatbot state when it cannot answer a question, and to then connect with human support, which then updated the knowledge base with their solution. This approach works pretty well after taking a few iterations to get right.

Companies that purchase large quantities of carbon credits typically follow one of two approaches: they either offset a fixed portion of their emissions (e.g. Delta Air Lines, easyJet, and Volkswagen Group) or offset residual emissions to meet specific emission targets (e.g. Shell, Eni, and Inpex). In contrast, companies that purchase fewer carbon credits typically only offer them to their customers during the checkout (e.g. Deutsche Lufthansa, Air France - KLM, Ryanair Holding, International Consolidated Airlines Group) or offset their business travel (e.g. Deutsche Lufthansa, International Consolidated Airlines Group).

“Whereas previously they had 100TWh contracted, we’re seeing nowhere near that kind of take-up under the new mechanism, and this is all happening in a climate where an increased amount of renewable sources are entering the system. So EDF is looking for replacement industrial off-takers for nuclear while this transition occurs.”

Mistral had already announced it was setting up its first dedicated AI training cluster in France. The company, which this week announced a fresh €1.7 billion ($1.99bn) funding round, had previously relied on infrastructure from Microsoft Azure and Google Cloud to train its models, but will now be working with Eclairion and French GPU cloud provider Scaleway to bring an initial 18,000 Nvidia GB200 GPUs online later this year.

This figure is true, as is the fact that much of Texas is in drought, but the usage covers the entirety of 2023 and 2024, suggesting data centers averaged a little over 600,000 gallons a day of water use. By contrast, Texas homes use 2.3b gallons of water every single day. Even the data quoted by the Austin Chronicle suggests data centers could use a comparatively modest 0.13b gallons a day in 2025. They might use a lot of water, but data centers are a tiny contributor to overall water use.

In a strategic reserve, a select few resources – usually fossil fuel plants close to retirement – are kept open in case their capacity is needed to fill gaps. These plants are put in a stand-by reserve and are not allowed to participate in electricity markets, instead getting paid by a central authority to simply be available and ensure they can be activated quickly during critical periods. Germany, for example, currently has a strategic reserves mechanism

In California, where record-breaking wildfires earlier this year collectively accounted for $37.5 billion in insured losses—roughly 70 percent of global insured disaster costs—insurance providers are adapting their models

The US, with its expanding population, relatively consistent economic growth and extensive domestic coal reserves, sees a cap on carbon emissions as a threat to its competitiveness, and hence to its global hegemony

Europe, meanwhile, has been making encouraging though patchy progress towards its Kyoto targets, driven mainly by a one-off switch from coal to gas – leaving it cradled in Russia’s increasingly rough embraces

The authors therefore suggest the government introduce a special battery grid fee, which would set a variable price for withdrawing or feeding electricity every quarter-hour, depending on the regional capacity of grids to transport electricity.

A security of supply report that systematically ignores large-scale storage facilities does not provide a viable basis for political decisions on power plant expansion or capacity markets

In this system, the electricity sector cannot balance out shortcomings in transport and buildings.

The “nameplate” power rating of a chip, often based on its thermal design power (TDP), reflects the maximum theoretical power it could draw under extreme, fully saturated workloads. These ratings are used to inform the electrical system design, but in practice, chips rarely operate near those levels. Reaching 100% of nameplate power across an entire rack or cluster is not only unusual, it's extremely difficult to sustain, and often not desirable given reliability and thermal constraints.

Load factor is the ratio of average demand to realized peak demand

As a result, all the "gains" from "powerful new models" come from burning more and more tokens. The cost-per-million-token number is no longer an accurate measure of the actual costs of generative AI, because it's much, much harder to tell how many tokens a reasoning model may burn, and it varies (as Theo Browne noted) from model to model

The "early days" argument hinges on obscurity and limited resources, something that generative AI does not get to whine about. Companies that make effectively no revenue can raise $500 million to do the same AI coding bullshit that everybody else does.In simpler terms, these companies are flush with cash, have all the attention and investment they could possibly need, and are still unable to create a product with a defined, meaningful, mass-market use case.

In one sense, explaining what happened to the telecom sector is very simple: the growth in capacity has vastly outstripped the growth in demand. In the five years since the 1996 bill became law, telecommunications companies poured more than $500 billion into laying fiber optic cable, adding new switches, and building wireless networks. So much long-distance capacity was added in North America, for example, that no more than two percent is currently being used

I was probably 9 or 10 when my father took me to a Horn & Hardart automat. For those too young to remember — who I hope are a large fraction of my readers — these were establishments in which a variety of sandwiches and other foods were displayed behind glass doors. You would serve yourself by putting coins into a slot, which would unlock the door and let you extract your egg salad sandwich or whatever.At the time (and at my age) it seemed wonderfully futuristic: Food service without people! In reality, of course, automats weren’t automated; each required a substantial staff to operate the kitchen and keep refilling those glass-doored compartments. And because automats weren’t all they pretended to be, they were eventually driven out of business by the rise of fast food.

Our study also determines that the minimum wage has no effect on employment in competitive markets, but it has a positive and very large effect in monopsonistic markets. If the minimum wage increases 100 percent, employment increases 69.7 percent more in monopsony markets than in competitive markets, where it only sees a tiny increase

Instead of appearing the ring N times, each item appears once, and takes advantage of fault-tolerant nodes rather than spreading over multiple nodes

Another troubling point is that big tech doesn’t just buy data centers and then use them, but in many cases pays a construction company to build them, fills them with GPUs and then leases them from a company that runs them, meaning that they don’t have to personally staff up and maintain them

And no, Microsoft talking about its Azure revenue in its last quarterly earnings for the first time is not the same thing, as it stopped explicitly stating their AI revenue in January (when it was $13 billion annualized).

Cursor is allowed to die, it will be unable to pay a chunk of Anthropic’s revenue — and yes, the revenue of other providers too. It will also bring into question whether it’s possible to build — putting aside any questions of profitability — a business of any kind offering services built on top of generative AI, and in turn bring into doubt the veracity of investing in this sector.