Cette section n'est pas encore traduite. Le contenu ci-dessous est affiché en anglais, tel qu'il a été publié. Lire la version anglaise.
AI news · updated 11 November 2026
Models, chips, deals, money and the people behind them, from 17 AI companies in the US, China, France and Canada. Short, plain English, and every story links to where it came from. Rumours and predictions are kept apart and labelled.
34
sourced stories
17
companies covered
5
AI bots tracked
11
rumours, theories and predictions, labelled
NVIDIA 
Why it matters: DSX POD ships Vera Rubin racks in containers with liquid cooling and rack power fitted and tested at the factory. NVIDIA says deployment at a prepared site drops from six months to as little as 14 days, so the site's power and cooling must be ready first.
ChipsData centres
Source: NVIDIA Newsroom (opens in a new tab) · Official post on X (opens in a new tab)
Google
Why it matters: DeepMind says its new vision-language-action model, built with Gemini 4 Argon, was tested on more than 200 tasks, finishes tasks 20% to 30% faster than the previous version and packed automotive parts with a 96% success rate in proof-of-concept tests at Mercedes-Benz. It is rolling out through partners including Agile Robots, Apptronik and Boston Dynamics. The figures are DeepMind's own; more robot training and inference means more GPU demand in data centres.
Models
OpenAI
Why it matters: CNBC confirmed OpenAI is now seeking up to US$20bn, down from about US$30bn, and the FT put the post-money valuation at US$1.25tn. The round is still open. Less fresh equity means more scrutiny of every gigawatt OpenAI has signed for.
Money
Source: CNBC (opens in a new tab)
Google
Why it matters: Google says it will support a shared AI training control in robots.txt, alongside Microsoft, OpenAI and Anthropic. It is based on an IETF draft. Website owners get one switch instead of one per crawler.
Policy
Amazon
Why it matters: Amazon repeated its outlook: "We continue to expect 2026 cash capital expenditures to be approximately $220 billion."
Data centres
Cohere 
Why it matters: The Canadian government will take an equity stake of up to C$1.6 billion alongside NVIDIA, Bell and Schwarz Group. Cohere commits at least C$1.7 billion to more computing infrastructure in Canada by 2030 and 1,000 new technical jobs.
DealsData centres
Google
Why it matters: Alphabet put $51.4bn into capex in Q3, about 35% of its technical infrastructure spend going to data centres. Google Cloud revenue grew 103% to $30.7bn.
Data centresDeals
Source: Alphabet (SEC exhibit 99.1) (opens in a new tab), The Motley Fool (call transcript) (opens in a new tab)
Microsoft
Why it matters: Microsoft expects Q2 capex above $50bn and about $170bn for the calendar year, including finance leases.
Data centres
Google
Why it matters: Oracle will offer Google TPUs on OCI from calendar 2027, with OCI providing the power, cooling and rack-scale direct liquid cooling. Rival clouds now share chips, and liquid cooling is the default.
ChipsDeals
Source: Oracle (opens in a new tab)
Google
Why it matters: Google's most capable model reads and writes up to about 1M tokens in one go, with low, medium and high thinking levels. It costs $2/$10 per 1M tokens until 30 November.
Models
Source: Google AI for Developers (opens in a new tab) · Official post on X (opens in a new tab)
Date the Gemini 4 Argon release notes were last updated. Google gives no launch date on that page.
Microsoft
Why it matters: Microsoft AI's image models MAI-Image-2.6 and 2.6-Flash moved to general availability, and Microsoft AI says a 2.7 preview ranks #3 on Arena's image-editing board.
Models
Alibaba 
Why it matters: Qwen-Audio-3.1-Realtime-Flash takes text and audio in and out, supports function calling, web search and voice cloning, and has a 262,144-token context window, says Alibaba Cloud's model list. It runs on Alibaba Cloud Model Studio.
Models
OpenAI
Why it matters: OpenAI says Terra matches GPT-6 Astra on Agents' Last Exam at a tenth of the price per token and finishes tasks over 3x faster.
Models
Source: OpenAI (opens in a new tab) · Official post on X (opens in a new tab)
Anthropic
Why it matters: Fortune reports $542–567bn of compute commitments and a plan, "Anthropic Direct", to lease dedicated data centres whose partners secure land, power and buildings.
DealsData centres
Source: Fortune (opens in a new tab) · Official post on X (opens in a new tab)
Moonshot
Why it matters: K3.1 Pro has a 1M-token context and costs $2 in and $10 out per 1M tokens on the Kimi API, a third below Kimi K3. Moonshot says full weights follow by 4 November 2026. Its Terminal-Bench 4.0 score of 65.7% is Moonshot's own claim.
Models
Source: Kimi (opens in a new tab), Kimi Forum (launch date) (opens in a new tab)
ByteDance
Why it matters: Volcano Engine released Seedance 2.5 4K for film and TV work, with 10-bit colour and cheap 480p draft previews. It is open only to selected enterprise customers for now. High-resolution video generation is one of the heaviest inference loads.
Models
Source: TechNode (opens in a new tab)
Google
Why it matters: A5X instances on NVIDIA Vera Rubin NVL72 reach general availability next month (November 2026) and use Google Cloud's fourth-generation liquid cooling. Google and NVIDIA are also exploring ways to fit more capacity within data centre power budgets.
ChipsData centres
Source: NVIDIA Blog (opens in a new tab)
Meta
Why it matters: Meta raised the bottom of its 2026 capex range, including finance-lease principal, from $130bn to $135bn.
Data centres
Source: Meta Platforms (SEC Form 8-K, exhibit 99.1) (opens in a new tab)
Anthropic
Why it matters: A research-preview app where Claude works as a proactive agent and checks with you before acting on your behalf.
Models
Source: Claude on X (opens in a new tab)
xAI
Why it matters: The Super Intelligence segment made $7.5bn of revenue in Q3. Musk said the limit on deploying AI is "more in power and cooling and infrastructure than it is in chips".
Data centresDeals
Source: TechCrunch (opens in a new tab)
Mistral 
Why it matters: Four Voxtral 3 models cover transcription in 50+ languages and text to speech, with a managed voice agent pipeline. Mistral offers EU, US and global hosted endpoints or self-hosting, so audio can stay on your own infrastructure. Transcription starts at $0.0035 per minute.
Models
Source: Mistral (opens in a new tab), Mistral Docs changelog (opens in a new tab)
Meta
Why it matters: Meta's next Vera Rubin racks run fully liquid-cooled at over 170 kW each. Starting or pausing a training job can swing tens of megawatts in seconds, so Meta uses batteries and power capping.
Data centresChips
Source: Engineering at Meta (opens in a new tab) · Official post on X (opens in a new tab)
MiniMax
Why it matters: MiniMax says M3.1 Flash is nearly three times faster and 70% cheaper on agent tasks than its predecessor, at $0.15 in and $0.60 out per 1M tokens. Weights for the 428 billion parameter model followed on 23 October. Speed and cost figures are MiniMax's own.
Models
Source: MiniMax (opens in a new tab), TechNode (opens in a new tab)
DeepSeek
Why it matters: DeepSeek says V4.1-Flash-1017 finishes tasks faster with shorter answers at unchanged API prices, from $0.15 in and $0.60 out per 1M tokens off-peak. The weights are on Hugging Face under the MIT licence. Benchmark gains are DeepSeek's own figures.
Models
Source: DeepSeek API Docs (opens in a new tab), SiliconANGLE (opens in a new tab)
DeepSeek
Why it matters: DeepSeek's Search and Fetch APIs give agents web results and pages as text. Calls made through its own harness, Claude Code or Codex were free until billing began on 1 November 2026, Beijing time. More agent traffic means more inference load on DeepSeek's capacity.
ModelsDeals
xAI
Why it matters: SpaceXSI says its new speech-to-speech model sets a record for resolving customer support calls, at the same API price as Grok Voice 2.0.
Models
Source: SpaceXSI (opens in a new tab) · Official post on X (opens in a new tab)
Meta
Why it matters: Muse Wonder and Muse Wonder Full are Meta's most capable models, built for agent work in law, finance and health, with Active Memory and Active Focus.
Models
Source: Meta (opens in a new tab) · Official post on X (opens in a new tab)
Anthropic
Why it matters: Anthropic's fastest model costs $0.10/$0.50 per 1M tokens for prompts up to 100k. Anthropic says it costs around 75% less to run on average.
Models
Source: Anthropic (opens in a new tab)
Mistral
Why it matters: Mistral previewed Large 4 on 6 October and has since released the weights. It is a mixture-of-experts model with 52 billion active parameters and a 1M-token context, priced at $0.70 in and $2.10 out per 1M tokens. It is also on Microsoft Foundry and NVIDIA-hosted endpoints.
Models
Source: Mistral (opens in a new tab), Mistral (open weights post) (opens in a new tab)
Amazon
Why it matters: The speech-to-speech model improves reasoning, instruction following and tool calls with lower latency, says AWS. It speaks seven languages, has a 256K context window and runs on Amazon Bedrock in four regions including Europe (Stockholm), at Nova 2 Sonic prices.
Models
Z.ai
Why it matters: AWS now serves GLM-5.3, a 753 billion parameter mixture-of-experts model with about 40 billion active and a 1M-token context, to eligible enterprise customers. It gives Chinese open-weight models a route into US cloud procurement.
ModelsDeals
Microsoft
Why it matters: MAI-Transcribe-2-Streaming handles 60 languages at $0.54 per hour of audio (introductory). MAI-Voice-2.1 speaks 23 languages in one voice.
Models
xAI
Why it matters: Grok 4.7 uses a larger base model than Grok 4.6, at the same price and speed: from $2 per 1M input and $6 per 1M output tokens.
Models
Source: SpaceXSI (opens in a new tab) · Official post on X (opens in a new tab)
Apple
Why it matters: Siri AI launched in beta in English with iOS 27, powered by Apple Foundation Models built with Google. Server features have daily usage limits, and Apple says expanded access will cost a fee in future. It is not initially available in the EU on iPhone.
Models
The assistants people use day to day, in no order of preference. Each latest update comes from the maker's own page, checked 1 November 2026.
xAI
Always-on AI teammates with a cloud computer of their own. They sign in to your apps and inboxes, finish multi-step work and come back for approval. Launched in beta on 11 August 2026.
Latest : Team Bots: Bots a whole team can share, working from the same files, apps and context. xAI news (opens in a new tab)
OpenAI
ChatGPT Work is OpenAI's agent for longer tasks across connected apps and files, with scheduled tasks and approvals. Launched 9 July 2026.
Latest : GPT-6.1 Terra, a model for visual reasoning and computer use, is now in ChatGPT Work and Codex on paid plans. OpenAI Help Center (opens in a new tab)
Google
Gemini can switch from chat to tasks and keep working while you are away, across Google apps and more than 40 connected apps.
Latest : Generate, preview and edit documents, PDFs, slides and spreadsheets directly in the Gemini app. Google blog (opens in a new tab)
Anthropic
Chat and Cowork are now one Claude. Longer work runs in a cloud environment Anthropic manages, so it keeps going after you close the laptop.
Latest : The unified chat and Cowork experience became generally available for Claude Enterprise. Claude by Anthropic (opens in a new tab)
Microsoft
Copilot in Microsoft 365, with Chat and Cowork on one Home screen and long-running tasks that ask before sensitive actions.
Latest : Learning Agent added, plus new Copilot usage metrics, in the release notes for 13 to 27 October. Microsoft Learn (opens in a new tab)
Our weekly video round-up of the AI race. It loads from YouTube only when you press play.
14 leaders and researchers who shape the models, chips and data centres. Roles are as stated on each linked page. Photos are freely licensed and credited on each card.
Chips
NVIDIA
Co-founder, president and CEO of NVIDIA.
Source: Wikipedia: Jensen Huang (opens in a new tab), accessed 1 November 2026
Chips
AMD
President and CEO of AMD since 2014.
Source: Wikipedia: Lisa Su (opens in a new tab), accessed 1 November 2026
Labs
OpenAI
CEO of OpenAI since 2019.
Source: Wikipedia: Sam Altman (opens in a new tab), accessed 1 November 2026
Labs 
Google DeepMind
Co-founder of Google DeepMind and CEO of Isomorphic Labs. Shared the 2024 Nobel Prize in Chemistry for protein structure prediction.
Source: Wikipedia: Demis Hassabis (opens in a new tab), accessed 1 November 2026
Labs
Anthropic
Co-founder and CEO of Anthropic, founded in 2021 with Daniela Amodei.
Source: Wikipedia: Dario Amodei (opens in a new tab), accessed 1 November 2026
Labs
Mistral AI
Co-founder and CEO of Mistral AI, Paris.
Source: Wikipedia: Arthur Mensch (opens in a new tab), accessed 1 November 2026
Labs
DeepSeek
Founder and CEO of DeepSeek and co-founder of the hedge fund High-Flyer.
Source: Wikipedia: Liang Wenfeng (opens in a new tab), accessed 1 November 2026
Labs
xAI, Tesla, SpaceX
CEO of Tesla and SpaceX.
Source: Wikipedia: Elon Musk (opens in a new tab), accessed 1 November 2026
Hyperscalers
Microsoft
Chairman and CEO of Microsoft, CEO since 2014.
Source: Wikipedia: Satya Nadella (opens in a new tab), accessed 1 November 2026
Hyperscalers
Microsoft AI
CEO of Microsoft AI. Co-founded DeepMind and Inflection AI.
Source: Wikipedia: Mustafa Suleyman (opens in a new tab), accessed 1 November 2026
Hyperscalers
Alphabet
CEO of Google since 2015 and of Alphabet since 2019.
Source: Wikipedia: Sundar Pichai (opens in a new tab), accessed 1 November 2026
Researchers
University of Toronto
Computer scientist known for work on neural networks. Nobel Prize laureate and University Professor Emeritus at the University of Toronto.
Source: Wikipedia: Geoffrey Hinton (opens in a new tab), accessed 1 November 2026
Researchers
Stanford University
Professor of computer science at Stanford. Created ImageNet, the dataset behind the 2010s advances in computer vision.
Source: Wikipedia: Fei-Fei Li (opens in a new tab), accessed 1 November 2026
Researchers
NYU and AMI Labs
Professor at NYU. Former Chief AI Scientist at Meta. Co-founded Advanced Machine Intelligence Labs in December 2025.
Source: Wikipedia: Yann LeCun (opens in a new tab), accessed 1 November 2026
Not fact. Everything in this section is a rumour, a theory or a forecast. Each card says where it circulates and sits beside what is actually known, with its own source. We do not endorse any of it, and none of it is investment advice.
Rumour (5)Speculation (2)Prediction (4)
The claim
Reported by The Information, relayed by Reuters.
Circulates in: Reuters: OpenAI plans to buy Google's AI chips, The Information reports (opens in a new tab)
What is known
The claim
Reported by The Information. Bloomberg.
Circulates in: The Next Web: Nvidia in talks to guarantee debt for Nscale's Anthropic data centre (opens in a new tab)
What is known
The claim
Reported by Bloomberg first. CNBC and Reuters from a source.
Circulates in: CNBC: Anthropic in talks for second $45 billion Nscale data center deal: source (opens in a new tab)
What is known
The claim
Reported by Elon Musk on the SpaceX results livestream, reported by TechCrunch.
Circulates in: TechCrunch: SpaceX's AI business is beating expectations (opens in a new tab)
What is known
The claim
NVIDIA is reported to be in talks to back the debt for Nscale's West Virginia data centre for Anthropic, taking over the site temporarily if Nscale defaults. Nothing is signed.
Circulates in: TechCrunch: Nvidia weighs playing the cosigner for Nscale's AI data center in West Virginia (opens in a new tab)
What is known
The claim
Sceptics argue that when chip makers and clouds invest in or guarantee the AI labs that buy their chips and capacity, demand and valuations can look stronger than they are, and losses could spread if AI revenue disappoints.
Circulates in: Axios: Nvidia reignites "circular" AI concerns as it weighs OpenAI financing guarantee (opens in a new tab), Bloomberg: AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other (opens in a new tab)
What is known
The claim
A theory that most of the internet is now bot activity and automated content. It began as a conspiracy theory about deliberate control and is now used loosely to describe the flood of generative AI content.
Circulates in: Wikipedia: Dead Internet theory (opens in a new tab)
What is known
The claim
Joachim Klement, head of market strategy at Panmure Liberum, says his core conviction is that the AI bubble bursts in 2027 or 2028. His year-end 2027 target for the S&P 500 is 5,000 points, which he put at 36% below the level at the time.
Circulates in: Bloomberg, via Yahoo Finance: AI Bubble Risks Worst S&P 500 Crash Since 2008, Strategist Says (opens in a new tab)
What is known
The claim
Capital Economics forecasts the S&P 500 at 6,500 by the end of 2027, about 16% lower, if air keeps leaking out of the AI trade or the bubble bursts. It also expects a final tech-led rally in 2026 first.
Circulates in: CNBC: Firm that correctly called the AI rally slowdown says here's where to hide out (opens in a new tab)
What is known
The claim
Gartner analyst Will Sommer predicts some US$3.5tn of AI infrastructure spending could be written off by the end of 2030, and that 80% of neoclouds will stop operating independently by 2030, bought or gone under.
Circulates in: The Telegraph: The tech 'death spiral' that could destroy the AI boom (opens in a new tab)
What is known
The claim
Lab leaders keep giving dates. Trackers record Sam Altman saying OpenAI will have an internal system it calls AGI by the end of 2026, Dario Amodei putting "powerful AI" in 2027 to 2028, and Demis Hassabis near 2030. There is no agreed definition of AGI.
Circulates in: AGI Watch: AGI timeline predictions, tracked (opens in a new tab)
What is known
Checked 1 November 2026. Rumour means an unconfirmed report of a specific event. Speculation means a theory or interpretation. Prediction means a forecast with a date or a number. We update or remove items when they are confirmed, denied or overtaken. Not investment advice.
Counted from the sourced stories on this page. Coverage, not importance.
| Company | Stories |
|---|---|
| 6 | |
| xAI | 3 |
| Microsoft | 3 |
| Meta | 3 |
| Anthropic | 3 |
| OpenAI | 2 |
| Amazon | 2 |
| DeepSeek | 2 |
| Mistral | 2 |
| NVIDIA | 1 |
| Alibaba | 1 |
| Moonshot | 1 |
| Z.ai | 1 |
| MiniMax | 1 |
| ByteDance | 1 |
| Apple | 1 |
| Cohere | 1 |
None of these is fact. See each card for where it circulates and what is known.
| Label | Items |
|---|---|
| Rumour | 5 |
| Speculation | 2 |
| Prediction | 4 |
New: AI models compared. Sort and filter every model by price, context and weights, compare any two, work out costs for your token volume, and read our AI agents hub.
Chinese models (DeepSeek, Qwen) and AI chips (Huawei Ascend, Alibaba T-Head) are covered in China: data centres and AI infrastructure.
Open: DeepSeek, Alibaba, Mistral, Z.ai, MiniMax, Cohere. Closed: Amazon, ByteDance, Apple. Kimi K3.1 Pro weights are due by 4 November. The original seven cards do not state this yet.
| Company | Model | In $/1M | Out $/1M | Context | Weights |
|---|---|---|---|---|---|
| Anthropic | Claude Haiku 5.5 | 0.1 | 0.5 | n/p | n/p |
| MiniMax | MiniMax M3.1 Flash | 0.15 | 0.6 | 1M | Open |
| DeepSeek | DeepSeek-V4.1-Flash-1017 | 0.3 | 1.2 | 1M | Open |
| Mistral | Mistral Large 4 | 0.7 | 2.1 | 1M | Open |
| OpenAI | GPT-6.1 Terra | 1 | 5 | n/p | n/p |
| Z.ai | GLM-5.3 | 1.4 | 4.4 | 1M | Open |
| Google | Gemini 4 Argon | 2 | 10 | 1.05M | n/p |
| xAI | Grok 4.7 | 2 | 6 | 500K | n/p |
| Alibaba | Qwen3.8-Max | 2 | 6 | 1M | Open |
| Moonshot | Kimi K3.1 Pro | 2 | 10 | 1.05M | Due |
| Meta | Muse Wonder | 4 | 20 | n/p | n/p |
| Microsoft | MAI-Voice-2.1 and MAI-Transcribe-2-Streaming | n/p | n/p | n/p | n/p |
| Amazon | Amazon Nova 2 Lite and Nova 2.5 Sonic | n/p | n/p | 1M | Closed |
| ByteDance | Seed2.1 Pro and Turbo | n/p | n/p | n/p | Closed |
| Apple | Apple Foundation Models 3 | n/p | n/p | n/p | Closed |
| Cohere | Command A+ | n/p | n/p | n/p | Open |
n/p = not published by the company on the pages we checked. Prices are list prices for standard requests and can change. Muse Wonder uses Cursor's listed price.
OpenAI
Fast model for computer use, visual reasoning and writing.
API price: $1 in / $5 out per 1M tokens
Speed and benchmark comparisons are OpenAI's own claims. we have not seen independent tests yet.
Sources: openai.com (opens in a new tab), x.com (opens in a new tab)
Google No launch date published
Google's most capable model for hard coding, agent, maths and science work.
API price: $2 in / $10 out per 1M tokens to 30 Nov 2026. $4 / $20 from 1 Dec
Features and prices are from Google's developer docs. Google gives no launch date there, and the price doubles on 1 December 2026.
Sources: ai.google.dev (opens in a new tab)
xAI
SpaceXSI's model for coding, agent tasks and knowledge work.
API price: $2 in / $6 out per 1M tokens (under 200k prompt tokens)
CursorBench scores are SpaceXSI's own published figures (manufacturer claim). Prices from SpaceXSI's API release notes.
Sources: x.ai (opens in a new tab), docs.x.ai (opens in a new tab)
Anthropic
Anthropic's fastest, cheapest model for high-volume, quick tasks.
API price: $0.10 in / $0.50 out per 1M tokens (prompts up to 100k)
OSWorld scores and the 75% saving are Anthropic's own figures (manufacturer claim). The saving depends on prompt size.
Sources: anthropic.com (opens in a new tab)
Meta
Meta's model for personal and professional agent tasks.
API price: Cursor's listed price: $4 in / $20 out per 1M tokens (Meta has not published one)
Meta published no price or benchmark table in its announcement. The price shown is Cursor's listed price, not Meta's.
Sources: about.fb.com (opens in a new tab), x.com (opens in a new tab)
Microsoft
Speech-to-text and text-to-speech for voice agents.
API price: Transcribe: $0.54 per audio hour (introductory to end of 2026). Voice-2.1: $22, Flash: $15 per 1M characters
Rankings and speed figures are Microsoft's own claims, citing Artificial Analysis for accuracy.
Sources: microsoft.ai (opens in a new tab)
DeepSeek
DeepSeek's fast, low-cost reasoning model for coding, agents and vision.
API price: $0.30 in / $1.20 out per 1M tokens at peak. half that off-peak
Benchmarks are DeepSeek's own (manufacturer claim). SiliconANGLE notes Claude Opus 5 still leads on most of DeepSeek's published tests.
Sources: api-docs.deepseek.com (opens in a new tab), huggingface.co (opens in a new tab), alibabacloud.com (opens in a new tab)
Alibaba
Alibaba's largest Qwen model, for coding, professional work and long agent tasks.
API price: $2 in / $6 out per 1M tokens
Arena rankings and benchmark scores are Alibaba's own (manufacturer claim).
Sources: alibabacloud.com (opens in a new tab), qwencloud.com (opens in a new tab), huggingface.co (opens in a new tab)
Mistral
Mistral's flagship multimodal model for coding, cyber defence and enterprise work.
API price: $0.70 in / $2.10 out per 1M tokens
Terminal-Bench gain (13 points to 41%) is Mistral's claim. It cites Artificial Analysis for its Coding Agent Index score of 48%.
Sources: mistral.ai (opens in a new tab), mistral.ai (opens in a new tab), docs.mistral.ai (opens in a new tab)
Amazon
Amazon's own models for everyday reasoning (Lite) and real-time voice agents (Sonic).
API price: Published on the Amazon Bedrock pricing page by region. Nova 2.5 Sonic costs the same as Nova 2 Sonic
Comparisons with rival models are Amazon's claims. Prices vary by region and tier, so check the Bedrock page.
Sources: aws.amazon.com (opens in a new tab), aws.amazon.com (opens in a new tab), aws.amazon.com (opens in a new tab)
Moonshot
Moonshot's most capable model, for coding, 3D creation, finance research and long agent tasks.
API price: $2 in / $10 out per 1M tokens. Highspeed $4 / $20
Scores are Moonshot's own (manufacturer claim), on an internal benchmark. Weights were not yet out on 31 October.
Sources: kimi.ai (opens in a new tab), kimi.ai (opens in a new tab), platform.kimi.ai (opens in a new tab)
Z.ai
Z.ai's flagship for coding, long agent tasks and security research.
API price: $1.40 in / $4.40 out per 1M tokens
Benchmarks are Z.ai's own (manufacturer claim). Z.ai says it still trails the top closed models on exploit benchmarks.
Sources: z.ai (opens in a new tab), docs.z.ai (opens in a new tab), aws.amazon.com (opens in a new tab)
MiniMax
MiniMax's fast, cheap model for agents, coding, computer use and security work.
API price: $0.15 in / $0.60 out per 1M tokens
Speed, cost and benchmark figures are MiniMax's own (manufacturer claim).
Sources: minimax.io (opens in a new tab), platform.minimax.io (opens in a new tab), technode.com (opens in a new tab)
ByteDance
ByteDance's model family behind the Doubao assistant, for agents, coding and video understanding.
API price: Not published in US dollars on the sources we checked
Rankings are ByteDance's own (manufacturer claim). No context window published.
Sources: seed.bytedance.com (opens in a new tab), seed.bytedance.com (opens in a new tab)
Apple
Apple's on-device and cloud models behind Siri AI and Apple Intelligence.
API price: No per-token API price. Server features have daily limits. Apple says a paid tier is coming
Quality results are Apple's own human preference tests (manufacturer claim). Announced 8 June 2026, released to users 14 September.
Sources: machinelearning.apple.com (opens in a new tab), apple.com (opens in a new tab)
Cohere
Cohere's enterprise model for agents, reasoning, vision and 48 languages.
API price: Not listed on Cohere's public pricing page
Capability claims are Cohere's own. No context window or token price published on the pages we checked.
Sources: docs.cohere.com (opens in a new tab), cohere.com (opens in a new tab)
Official uploads from each company's YouTube channel. Nothing loads from YouTube until you press play, and then it uses youtube-nocookie.com. We did not find an official SpaceXSI (xAI) YouTube upload; see its posts on X below.
Quarterly capex for the biggest AI and cloud spenders, with their guidance, is on Market intelligence. Rack power and cooling figures are in chip power and IT equipment. Listing a company here is not a recommendation.
The latest news from the big AI companies, the people behind them and what it means for power and cooling, each story with a link to where it came from.
Every new AI model has to be trained and run on thousands of chips in data centres, so AI plans turn into building and electricity plans.
Why it matters The IEA calls AI the most important reason data centre electricity use is set to more than double by 2030.
Source: IEA, Energy and AI (2025)
Not necessarily. They are things people are saying, shown next to the known facts and clearly labelled Rumour, Speculation or Prediction.
Why it matters Treat them like a weather forecast, not a news report. Nothing here is investment advice.
Regularly. The date at the top of the page shows the last update.
Yes. This Week in AI is a short weekly video that sums up the main stories. It is on the page and on our YouTube channel.
