Intelligence. Capitalised The definitive AI authority for the global executive. Mapping the strategy and economics of Artificial Intelligence.
Barcelona | Global
Joined April 2026
- Tweets821
- Following116
- Followers49
- Likes109
tokenando retweeted
Whenever someone asks me whether AI will make it easier for hackers to move money out of our bank accounts, my answer is always the same:
the problem is usually the human in the loop.
The Compute by @TokenandoAI Issue XXV is out:
tokenandoai.substack.com/p/t…
🤖 Made with AI
BlackRock chairman and chief executive Larry Fink warned in a Bloomberg video interview published October 21 that a slow data-centre buildout carries its own inflation risk. Fink argued that insufficient AI infrastructure investment could constrain the productivity gains needed to offset inflationary pressures elsewhere in the economy. BlackRock manages approximately $11.5tn in assets and has significant exposure to infrastructure investment through its alternatives and real assets platforms.
Fink's position sits in direct tension with signals from central bankers gathered at the IMF and World Bank annual meetings in Bangkok in mid-October 2026. Federal Reserve Chair Kevin Warsh described AI investment as a demand-side shock arriving before the supply-side productivity offset, and warned the transition could be "very bumpy." ECB policymaker Martins Kazaks told Reuters there is "a global fight for money" and that AI investment is "one of the key factors driving up the cost of money."
The 2 positions are not simply a disagreement about pace. Warsh's concern is that building too fast generates demand-side inflation before productivity gains materialise. Fink's concern is that building too slowly would produce a productivity shortfall with its own inflationary consequences. Both arguments are being made simultaneously by credible, senior figures, and the uncertainty between them is a live variable for any organisation making long-duration infrastructure commitments.
🤖 made with AI
Full details via Bloomberg / bloomberg.com/news/videos/20…
Ericsson ConsumerLab published a report in October 2026 based on a survey of more than 43,000 smartphone users aged 15 to 69 across 27 markets, examining how consumer AI usage is shaping future network needs. The UAE findings stand out: 10% of respondents already use agentic AI for tasks they previously performed themselves, double the global figure of approximately 4 to 5%.
By 2030, 63% of UAE smartphone users are expected to delegate everyday tasks to AI agents, against a global projection of 38%. The share of AI super users, defined as those making heavy use of multimodal AI, is projected to grow from 18% to approximately 46% of UAE smartphone owners by 2030. For this group, AI response time is 2.4 times more critical to network satisfaction than for the average user. Separately, 54% of daily AI users surveyed expressed strong concern about fake AI assistants acting in their name; Ericsson's data identifies communication service providers as the most trusted verifiers of AI agent identity.
e& UAE announced in August 2026 that it is embedding agentic AI directly into its Business Pro fixed internet and enterprise connectivity plans, pairing the capability with Sovereign AI Compute providing UAE-based GPU capacity. Ericsson Gulf Region president Petra Schirren has noted the UAE's adoption trajectory is running ahead of global averages.
🤖 made with AI
Full details via Ericsson / ericsson.com/en/press-releas…
AT&T published details on October 21 of its partnership with OpenAI to build AI-enabled workflows across its in-house legal department. The collaboration builds on LegalEdge, an AI-first centre of expertise launched in January 2026 that handled nearly 100 matters in its first 6 months and has grown from 24 to approximately 35 people. On October 8, AT&T was formally designated a flagship in-house legal design partner for OpenAI.
The planned system includes a natural language intake front door that classifies and prioritises incoming legal requests, connects them with relevant precedent and past work, and provides the responsible lawyer with historical context and recommended first steps. AI agents may answer frequently asked questions directly; for complex matters they gather and synthesise information for a lawyer to act on. AT&T general counsel David McAtee has indicated the firm's approximately 30 national law firms could reduce to around 18 to 20 over the next 2 years as more work moves in-house, though no firm target has been set.
Chief compliance officer Bill Ryan has stated that success will be measured by quality, knowledge accessibility, and legal capacity rather than by technology deployment volume.
🤖 made with AI
Full details via OpenAI / openai.com/index/att-legal/
Jas Khaira, who leads Blackstone's AI investment unit BXN1, told TechCrunch Disrupt in October 2026 that concerns about systemic risk from AI infrastructure debt are overstated. Blackstone manages $1.3tn in assets and describes itself as the largest owner of compute outside the hyperscalers, with positions spanning data centres, neocloud lenders, and direct equity in frontier AI companies.
Khaira's argument centres on the creditworthiness of the end customer rather than the intermediary. Because most data centres are built for hyperscalers, he contends the underlying debt is backed by the most creditworthy companies in the world. An internal review of BCRED, Blackstone's private credit fund, found under 5% of investments face substantial AI-related headwinds. AI-focused leveraged finance issuance reached $88bn in 2026, up from approximately $20bn the prior year.
The market is sending a different signal in some corners. CoreWeave's bonds fell below 80 cents on the dollar in October 2026, with yields approaching 15%. Firmus, an Australian GPU provider to which Blackstone co-led a $10bn credit facility alongside Coatue, abandoned its planned ASX listing at a target valuation of approximately $30bn. Khaira acknowledged rising rates and wider credit spreads, and described AI ROI across Blackstone's portfolio companies as "real but uneven," following a "sawtooth-type pattern" where compute investment precedes utilisation. Blackstone also operates a $1.5bn joint venture called Ode with Anthropic and Hellman & Friedman, supplying forward-deployed engineers to its 280 portfolio companies.
🤖 made with AI
Full details via TechCrunch / techcrunch.com/2026/10/21/bl…
Google Cloud announced on October 21 that OpenAI's frontier models are joining its Model Garden, making it the only hyperscale provider offering all 3 leading closed-source model families on a single platform: Gemini, Claude, and GPT.
3 OpenAI models are available in private preview for select enterprise customers. GPT-6 Astra is positioned for demanding tasks including coding and scientific research. GPT-6.1 Sol targets agentic and professional work at lower cost. GPT-6 Luna is designed for high-volume everyday tasks. All usage is billed through a single Google Cloud invoice, counts toward existing cloud spending commitments and Flexible Savings Plans, and carries Google's data governance guarantees, including a commitment that customer data will not be used to train foundation models.
OpenAI's Codex CLI and Desktop workflows can already connect to the Gemini platform for inference, with broader availability in Gemini for Business and Google Antigravity to follow. Microsoft remains OpenAI's primary cloud partner, but OpenAI models are now available across all 3 major US cloud providers. General availability is described as coming within weeks.
🤖 made with AI
Full details via Google Cloud Blog / cloud.google.com/blog/produc…
Demis Hassabis now expects artificial general intelligence within 2 to 3 years. The Google DeepMind chair and Alphabet chief scientist told the Wall Street Journal on October 21 that he is "updating nearer and nearer," compressing a timeline that stood at 3 to 5 years earlier in 2026 and at 5 to 10 years when he spoke at Davos in January of that year.
Hassabis defines AGI as a system capable of the full range of human cognitive abilities, distinct from performance on narrow benchmarks. He pointed to Isomorphic Labs, his drug-discovery venture that raised $2.1bn this year, as a near-term illustration of what AI-driven science can already do. He has also acknowledged that his public statements are designed in part to create urgency among governments and the public, meaning the timeline functions as both a forecast and a policy instrument.
The convergence with Anthropic is notable. Chief executive Dario Amodei made a near-identical 2-to-3-year prediction at Davos in January 2025, meaning the two leading frontier labs now share a public window. Hassabis has proposed the Einstein Test as a concrete AGI benchmark; no published result shows any system has passed it, giving organisations an observable threshold to monitor alongside the public timeline debate.
🤖 made with AI
Full details via The Verge / theverge.com/ai-artificial-i…
📰 The Tokenando Daily / Thursday October 22, 2026
/ Demis Hassabis compresses his AGI timeline to 2 to 3 years, converging with Anthropic's Dario Amodei on the same window
/ Google Cloud adds OpenAI's GPT-6 models to its platform, becoming the only hyperscaler offering Gemini, Claude, and GPT under one invoice
/ Blackstone's Jas Khaira argues AI infrastructure debt fears are overstated, even as CoreWeave bonds trade below 80 cents on the dollar
/ AT&T and OpenAI formalise a legal AI partnership, with AT&T signalling its external law firm panel could shrink from 30 firms to around 18 to 20
/ Ericsson ConsumerLab finds 63% of UAE smartphone users expect to delegate daily tasks to AI agents by 2030, double the global projection
/ BlackRock's Larry Fink warns that a slow data-centre buildout carries its own inflation risk, in direct tension with central bank caution on AI-driven demand
🤖 made with AI
✅ reviewed by the Tokenando team
🔗 links and details in following posts
🤖 Made with AI
A leaked S-1 filing reported by The Wall Street Journal reveals that Anthropic is experiencing structural tensions in its relationships with Amazon Web Services, Google Cloud, and Microsoft, which the filing describes as increasingly viewing Anthropic as a competitor rather than a customer. Anthropic's total projected cloud spending could exceed $750bn, including deals still under negotiation, yet the company reports it cannot secure sufficient compute capacity from the hyperscalers and is supplementing with leased data centres and third-party operators it describes as having "limited operating history."
The company reported $16.3bn in revenue and an $80.7bn net loss in the first half of 2026, with 2 unnamed customers accounting for approximately one-fifth of sales. Third-quarter 2026 revenue reached $16.8bn, up 45% from the prior quarter. AWS and Google routed 47% of Anthropic's 2025 revenue through their cloud marketplaces, up from 32% in 2024 and 11% in 2023, while also holding equity stakes, supplying compute, and competing directly with Claude through their own model offerings. Anthropic paid $351m in distribution fees to AWS and Google and held $54.6bn in non-cancellable hosting commitments at end-2025.
A separate dispute with OpenAI concerns Anthropic's practice of booking gross marketplace revenue, which OpenAI has argued inflates reported figures by billions of dollars. Anthropic is targeting a November 2026 IPO. The Nscale $45bn compute deal is cited as Anthropic's primary attempt to diversify supply away from the hyperscalers.
🤖 made with AI
Full details via TechCrunch / techcrunch.com/2026/10/20/an…
On October 20, 2026, 2 companies released production-ready voice AI systems within hours of each other, compressing what had been a relatively uncrowded market into a contested one within a single day.
Mistral AI released Voxtral 3, a complete voice agent system comprising 3 speech-to-text models, 2 text-to-speech models, and a voice agent SDK built on the LiveKit framework. The system combines transcription, reasoning, tool calling, and speech generation into a single real-time streaming pipeline, with each turn completing in under 1 second. Transcription supports more than 50 languages; text-to-speech covers more than 35 languages and 50 dialects with emotional adaptation. Pricing starts at $0.0035 per minute for batch transcription and $0.04 per minute for the full voice agent stack using Mistral Small as the reasoning model. The system is available through Mistral's EU-hosted API or for self-deployment on private infrastructure. Mistral's own figures show sub-500ms transcription latency and a 4 to 4.1% word error rate on the FLEURS 45 multilingual benchmark; no independent benchmarking had been published at launch.
ElevenLabs released Eleven Realtime on the same day, offering 2 variants: v1 Mini, optimised for natural conversation at lower cost, and v1 Max, designed for complex tasks including customer support and multi-step workflows with tool calling. Both support more than 90 languages and can draw on more than 17,000 voices. Independent evaluation by Artificial Analysis found that v1 Max answered 98.8% of questions correctly on the Big Bench Audio benchmark with a 0.57-second time to first audio; v1 Mini scored 75.9% with a 0.62-second response time. On tool-use tasks, v1 Max completed 87.1% successfully against 60.3% for v1 Mini. Both models are compatible with OpenAI's Realtime API specification, allowing developers to switch by changing the endpoint and model name in existing code.
🤖 made with AI
Full details via Mistral AI / mistral.ai/news/voxtral-3/
Full details via ElevenLabs / elevenlabs.io/blog/realtime
Anthropic has been circulating a draft federal AI safety bill, tentatively called the Artificial Intelligence Evaluation, Governance, and Incident Safety Act (the AEGIS Act), to US lawmakers since at least October 2026, as first reported by Politico on October 16, 2026. The bill would allow the Commerce Department to seek a court order to pause an AI model's release only when it poses a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or causing more than $1bn in property damage, arising from a single incident.
The triggering scenarios are limited to 3 categories: the creation or release of a biological weapon, a model evading its developer's controls, or an AI-enabled cyberattack. Independent evaluators registered with the Commerce Department would periodically assess frontier models; labs could pay these evaluators provided payment is not contingent on the evaluation's outcome. If an evaluator identifies imminent risk, they must notify Commerce within 72 hours, after which Commerce would present evidence to a federal judge in the District Court for the District of Columbia. The bill explicitly prohibits any other federal department, agency, or officer from ordering a halt to a model's release.
The bill's thresholds mirror those in California's Senate Bill 53, which Anthropic supported in 2025. The legislation has no formal congressional sponsors as of the reporting date. Anthropic's federal affairs head Anthony Cimino has publicly defended the thresholds on X, citing the California law precedent.
🤖 made with AI
Full details via Politico / politico.com/news/2026/10/16…
SpaceX reported third-quarter 2026 results on October 20, 2026, with revenue of $14.58bn against analyst expectations of $12.72bn, and net income of $4.4bn reversing a loss of $1.05bn in the same period a year earlier. The company's AI and cloud division, now branded "Super Intelligence," generated $7.56bn in revenue, up ninefold from $813m a year earlier. Capital expenditure reached $19.6bn in the quarter, with more than $16.4bn committed to the AI business.
Revenue drivers included cloud services agreements with Google and Anthropic, the acquisition of coding platform Cursor for $60bn, and growth in Grok adoption. SpaceX guided first-quarter 2027 revenue of $33bn to $37bn, against analyst consensus of $22.2bn. The company is also developing Terafab, a chip fabrication facility targeting volume production in Q3 2028 with Intel and Tesla as partners.
On the earnings call, chief executive Elon Musk confirmed that both SpaceX and Tesla have formed special committees of independent directors to discuss a potential merger at shareholder request, and stated that he personally supports the combination. A merger would require approval from 50% of all Tesla shareholders. Mr Musk holds more than 82% of voting power at SpaceX and approximately 20% at Tesla. The combined market capitalisation of the 2 companies at the time of the announcement was approximately $3.5tn.
🤖 made with AI
Full details via CNBC / cnbc.com/2026/10/20/spacex-s…
Nvidia is in talks to invest hundreds of millions of dollars in Firmus Grid, an Australian AI data centre operator that withdrew its planned ASX listing on October 9, 2026, according to The Information. The IPO had sought to raise approximately $5bn at a valuation of roughly $30bn, nearly triple Firmus's $10.5bn valuation from August 2026. Prospective buyers cited the gap between a proposed $44bn market capitalisation and $50m in annual revenue, alongside concerns about existing shareholders' ability to sell large portions of stock immediately after listing.
Firmus counts OpenAI and Meta among its customers, holds a $10bn debt facility led by Blackstone and Coatue, and has 46 MW of operational capacity against 865 MW under development across Australia, Singapore, Malaysia, and Indonesia. Nvidia already holds a 7.2% stake in the company following 3 prior investment rounds. The valuation trajectory from $1.85bn in September 2025 to a proposed $44bn in October 2026 covers 13 months.
Morgan Stanley has estimated that AI infrastructure could require approximately $1.5tn in external financing through 2028. Firmus is separately pursuing a private round of $2bn to $3bn, with the structure of any Nvidia investment still under discussion.
🤖 made with AI
Full details via The Information / theinformation.com/articles/…
DeepSeek released DeepInfer on October 20, 2026, an open-source inference framework under the Apache 2.0 licence that powers its V4.1-Flash model in production. The company reports that a single cluster running V4.1-Flash serves approximately 155 trillion tokens per day, generating revenue equivalent to $38 per GPU per hour at grid electricity prices. That figure is the first time DeepSeek has attached a production economics number to its inference stack.
The release includes optimised computing kernels, speculative decoding, distributed inference support, and a companion tokenisation library called deepseek-recipe. From day one, DeepInfer runs on 2 hardware platforms: Nvidia's SM100 (Blackwell) GPUs and Huawei's Ascend 950 NPUs. The scope is deliberately bounded: the release covers offline inference and core serving features but excludes distributed KV cache management, prefill-decode separation infrastructure, and load balancing across replicas.
For context on the cost differential the release is designed to widen: DeepSeek V4.1-Flash is priced at $0.035 per million input tokens against OpenAI GPT-6 Astra at $10 per million input tokens. The release came 1 day after the US government ordered civilian federal agencies to remove DeepSeek products from their systems. Open-weight models already account for 23% of spending on Vercel's AI Gateway; independent benchmarking of DeepInfer's performance claims has not yet been conducted.
🤖 made with AI
Full details via The Next Web / thenextweb.com/news/deepseek…
📰 The Tokenando Daily / Wednesday October 21, 2026
/ DeepSeek open-sources DeepInfer, the inference engine behind its V4.1-Flash model, reporting $38 per GPU per hour in production revenue
/ Nvidia in talks to invest hundreds of millions in Firmus Grid after the Australian data centre operator's $44bn IPO collapsed for insufficient demand
/ SpaceX AI division posts $7.56bn in quarterly revenue, up ninefold year on year, as Elon Musk confirms Tesla merger talks are under way
/ Anthropic circulates draft AEGIS Act to US lawmakers, proposing federal AI safety intervention only when more than 50 deaths or $1bn in damages are at stake
/ Mistral launches Voxtral 3, a full voice agent stack covering 50-plus languages, with EU-hosted API and on-premise deployment options
/ ElevenLabs releases Eleven Realtime, 2 speech-to-speech models with OpenAI API compatibility and independent benchmark scores covering 90-plus languages
/ Anthropic's leaked IPO prospectus reveals AWS and Google routed 47% of its 2025 revenue while simultaneously competing with Claude across enterprise markets
🤖 made with AI
✅ reviewed by the Tokenando team
🔗 links and details in following posts
🤖 Made with AI
The Aachen Regional Court issued a preliminary injunction on October 18, 2026 barring Google from spreading false statements about a physician through its AI Mode search feature, explicitly citing the reasoning of a May 2026 Munich Regional Court ruling. In the Aachen case, Google's AI had attributed criminal allegations involving a different former university professor to the physician, fabricating a detailed legal aftermath including invented proceedings. The doctor notified Google in late August 2026; Google declined to act in mid-September. The Aachen court acted on an emergency basis without hearing Google first, and set fines of up to €250,000 per violation.
The Munich ruling, handed down on May 28, 2026, classified AI Overviews as Google's own statements rather than neutral aggregation of third-party content, making Google directly liable for their accuracy. Google announced an appeal of the Munich ruling in June 2026 but subsequently settled, meaning no higher court has yet ruled on the underlying legal theory. The Aachen court described AI Overviews as producing "independent, new, and substantive statements based on an evaluation and combination of content from various third-party websites."
Parallel rulings have been noted in Frankfurt, Hamburg and Hamm. The distinction between the Munich and Berlin rulings is legally significant: Munich found direct liability for evaluative AI summaries, while Berlin found no liability for AI texts that remained within the underlying search results.
🤖 made with AI
Full details via The Decoder / the-decoder.com/two-german-c…
Emirates NBD and OpenAI announced on October 19, 2026 that the Dubai-based banking group would become the first bank in the MENA region to deploy OpenAI's GPT models using in-country inference infrastructure in the UAE. The arrangement covers 4 areas: customer engagement across digital channels, risk and compliance analytics, automation of knowledge-intensive workflows, and employee productivity.
Emirates NBD will connect the models to its internal systems, APIs and enterprise data. The bank has total assets of AED 1.3tn ($354bn), operates in 13 countries, serves more than 25 million active customers, and reports that 97% of its transactions and requests already occur outside branches. OpenAI switched on inference residency in the UAE in August 2026, making it the first MENA market where prompts, files and conversations are processed only on GPUs within the country.
No financial terms were disclosed, and no deployment launch date has been announced.
🤖 made with AI
Full details via Emirates NBD / emiratesnbd.com/en/media-cen…
CoreWeave's bonds fell into distressed territory during the week of October 13, 2026, with yields on its 2031 bonds reaching approximately 15 to 16% and 5-year credit default swap spreads exceeding 1,000 basis points. Market pricing implied a roughly 63% probability of default within 5 years. The company carries approximately $50bn in debt and lease obligations against first-half 2026 revenue of $4.6bn and a net loss of $1.4bn over the same period.
CoreWeave's shares fell more than 20% over the period, including a single-day decline of 15.5% on October 15. Co-founder Brannin McBee told the Financial Times that the company was seeking to tap insurance company capital, describing the sector's access to funds as "unlimited." Chief executive Mike Intrator issued an internal memo dismissing the market moves as irrational, citing a backlog exceeding $100bn, approximately two-thirds of which is with investment-grade counterparties including Microsoft, Meta, IBM and Caterpillar.
Gartner published a forecast the same week estimating that 80% of neocloud providers could cease independent operations by 2030 and that more than $3.5tn of AI spending could face impairment at current investment rates. Carlyle has estimated that private credit could be called on to finance approximately $1tn of the AI compute build-out. An Nvidia backstop arrangement, under which Nvidia will purchase unsold CoreWeave capacity through April 2032, is a structural feature of the market that has received limited attention in coverage of the current stress.
🤖 made with AI
Full details via The Verge / theverge.com/ai-artificial-i…
The US State Department designated 5 Chinese entities on October 19, 2026 under the Protecting American Intellectual Property Act, in connection with the theft of AI trade secrets from Google. The action followed the January 2026 federal jury conviction of Linwei Ding, a former Google software engineer, on 7 counts of theft of trade secrets.
Ding copied thousands of pages of confidential information relating to Google's AI infrastructure, including its Tensor Processing Unit systems, GPU clusters, networking technology and the software coordinating large chip arrays, between approximately May 2022 and April 2023. He founded Shanghai Zhisuan Technology in 2023 and applied to MiraclePlus, a Beijing-based venture capital firm, citing his access to Google's systems. The 5 designated entities are Shanghai Zhisuan Technology, MiraclePlus, and 3 affiliated investment partnerships.
Sanctions include asset blocking, restrictions on US financial institution lending above $10m in any 12-month period, and prohibitions on US persons investing in the entities' equity or debt. A wind-down general licence runs until January 19, 2027. A judge had partially reversed the economic espionage counts at trial, finding insufficient evidence that Ding knew his actions would benefit the Chinese government, but that finding did not prevent the State Department from proceeding with entity-level designations.
🤖 made with AI
Full details via US Department of State / state.gov/releases/office-of…
Anthropic is in advanced talks with Nscale, the Nvidia-backed UK data centre operator, for a second cloud computing arrangement worth approximately $45bn, according to Bloomberg reporting on October 19, 2026. The proposed deal would cover a new facility at Nscale's Monarch campus in West Virginia, built around Nvidia Vera Rubin chips and providing roughly 460 megawatts of capacity.
The October talks follow an existing agreement signed in August 2026, under which Anthropic committed to spend approximately $44.6bn over 6 years renting compute from Nscale's first West Virginia site. If the current negotiations conclude, total committed spend between the two companies would approach $90bn. Anthropic also holds a right of first refusal on a third planned building at the Monarch campus.
No contract has been signed. Bloomberg's account rests on unnamed people familiar with the negotiations, and neither company has issued a public statement. A separate Reuters report noted that Microsoft withdrew from an earlier Nscale anchor deal after power connection delays pushed that site's expected supply date into the 2030s. Nscale is separately preparing for a stock market listing targeting a valuation of approximately $30bn.
🤖 made with AI
Full details via Bloomberg Law / news.bloomberglaw.com/capita…
