tokenando

Intelligence. Capitalised
The definitive AI authority for the global executive.
Mapping the strategy and economics of Artificial Intelligence.

[Ziad Matar @wirelessly]

When the US starts doubting the effective execution of export controls, that they have been circumvented or are at risk of circumvention... the world should expect tighter restrictions!

[Ziad Matar @wirelessly]

A residual $17 MUSD financial interest in a China JV, disclosed in a 2025 SEC filing, was sufficient to trigger concern despite SK Telecom's minimal direct China operations. A very low threshold for what constitutes a problematic China connection in the current tense environment.

[tokenando @TokenandoAI]

2 studies published in "Nature" on June 17 and 18, 2026 showed specialised AI systems matching or exceeding physician performance on specific clinical tasks. MIRA, a German system built on GPT-4o and o1-preview, achieved diagnostic accuracy of 87.8% across more than 500 emergency department cases, against 78.1% for a panel of 6 physicians. Google's AMIE system, built on Gemini 1.5 Flash, produced treatment plans that aligned more closely with clinical guidelines than those of 21 primary care physicians across 100 multi-visit case scenarios.

A supplementary finding in the AMIE paper showed that when the same specialised setup was applied to the newer Gemini 2.5 Flash, the performance advantage over the base model almost disappeared. A separate study in "Nature Medicine" by NYU Langone Health found that general-purpose frontier models, including GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6, outperformed 2 commercially deployed specialised clinical AI tools across all 3 evaluation categories.

Both MIRA and AMIE were built on models already superseded at the time of publication.

Full details via The Decoder / https://t.co/qeXDljvkSB

🤖 made with AI

[tokenando @TokenandoAI]

Meta Platforms has signed contracts to purchase approximately 1.6GW of AI computing capacity from data centre developer Crusoe across 2 sites: Childress, Texas, and Warrenton, Missouri. Financial terms and delivery timelines were not disclosed.

Crusoe's existing contracted clients include Oracle, Microsoft, and Google. Its total contracted capacity now approaches 4.9GW across 5 US campuses, with a development pipeline exceeding 40GW. The company's 1.2GW Abilene, Texas campus, built for Oracle and used by OpenAI, has 2 buildings operational with 6 more under construction.

Crusoe raised $1.37bn in a Series E round at a valuation above $10bn, co-led by Valor Equity Partners and Mubadala Capital. Meta has separately pledged to spend at least $600bn on AI infrastructure over the next few years and is developing a nearly 4,000-acre campus in Louisiana targeting up to 5GW of capacity.

Full details via Bloomberg / https://t.co/I6UccmiV7I

🤖 made with AI

[tokenando @TokenandoAI]

The Federal Energy Regulatory Commission voted unanimously on June 18, 2026 to direct 6 regional grid operators, covering approximately 200 million Americans, to process AI data centre power connections within 90 days. The current interconnection process can take years.

Grid operators have 30 days to report on available generating capacity and 60 days to defend or revise existing electricity tariffs. Data centres will bear the full cost of any grid upgrades required for their connection. Hyperscalers may also be required to accept curtailment during periods of high grid stress.

FERC chair Laura Swett described the action as historic. Electricity demand from data centres is projected to nearly triple through 2035. The order follows a directive from Energy Secretary Chris Wright in October 2025 calling for expedited reviews, and comes amid political pressure as data centre construction has contributed to rising consumer electricity bills ahead of November mid-term elections.

Full details via Bloomberg / https://t.co/GxVHzy1wVF

🤖 made with AI

[tokenando @TokenandoAI]

SK Telecom has been identified as the South Korean company whose inclusion in Anthropic's Project Glasswing programme triggered US government concern and the export control directive issued on June 12, 2026. The concern centred on SK Telecom's historical relationship with China Unicom, including a 2004 joint venture, a $1bn bond investment in 2006, and a residual $17m financial interest disclosed in a 2025 SEC filing, despite SK Telecom having divested the majority of its stake in 2009.

Anthropic expanded Project Glasswing to approximately 150 institutions across 15 countries on June 2, 2026, including SK Telecom alongside Samsung Electronics and the Korea Internet and Security Agency. The White House ordered Anthropic to revoke SK Telecom's access; Anthropic complied immediately. A separate jailbreak allegation from Amazon compounded the crisis, resulting in the Commerce Department directing Anthropic to restrict Fable 5 and Mythos 5 to US citizens only.

Bloomberg subsequently reported that select early preview users have retained access to a Mythos preview environment despite the broader shutdown, which cut off commercial customers across the EU, UK, Canada, and India. Anthropic and the White House remained in active negotiation over restoration terms as of June 19, 2026.

Full details via Wired / https://t.co/LNm3dha3yz

🤖 made with AI

[tokenando @TokenandoAI]

US Commerce Secretary Howard Lutnick has raised concerns directly with ASML executives, alleging that one of the company's EUV lithography machines may have reached China in violation of export restrictions. ASML denied the claim, stating it has never shipped an EUV system to China and circulated documents in Washington to support that position.

The episode follows months of escalating pressure. The Dutch government has already revoked licences for older deep ultraviolet machines. A bipartisan group of US lawmakers introduced the MATCH Act in April 2026, which would extend restrictions to all DUV immersion tools and ban servicing of equipment at SMIC, Hua Hong, and CXMT.

China accounted for 33% of ASML's sales in 2025 and is projected to fall to around 20% in 2026 before any further restrictions take effect. A servicing ban, even without new sales restrictions, would degrade the performance of machines already installed in China.

Full details via Bloomberg / https://t.co/Nobe2lvLQo

🤖 made with AI

[tokenando @TokenandoAI]

📰 The Daily brief / Friday June 19, 2026

/ US Commerce Secretary Howard Lutnick raises concerns with ASML that an EUV lithography machine may have reached China in breach of export restrictions, as the MATCH Act threatens further curbs on Dutch chip tool sales

/ SK Telecom's historical ties to China Unicom identified as the trigger for the US government order that forced Anthropic to revoke Mythos access and issue a broader export control directive on June 12, 2026

/ Select early preview users retain access to Anthropic's Mythos model despite the Commerce Department directive that cut off all commercial customers across the EU, UK, and Canada

/ FERC votes unanimously to require 6 regional grid operators to process AI data centre power connections within 90 days, against a current process that can take years

/ Meta signs contracts for 1.6GW of AI computing capacity with data centre developer Crusoe across sites in Texas and Missouri, adding to a portfolio that already includes Oracle, Microsoft, and Google

/ 2 studies in "Nature" show specialised AI systems matching or exceeding physician accuracy, while a supplementary finding reveals the performance advantage nearly disappears when the same setup runs on a newer base model

🤖 made with AI
✅ reviewed by the Tokenando team
🔗 links and details in following posts

[Ziad Matar @wirelessly]

Straight off the press…

[tokenando @TokenandoAI]

On the evening of June 12, the most powerful model Anthropic had ever released was switched off. A US Commerce Department letter ordered it to block every foreign national from Fable 5 and Mythos 5. Nationality cannot be verified at the interface, so both went dark for everyone, worldwide.

That same weekend, Satya Nadella made the case for owning your AI rather than only renting it. He calls it token capital. The companies that had not treated AI as a strategic asset learned the cost on Friday.

Issue VII of The Compute is out.

https://t.co/qiFK4cUtkS

[tokenando @TokenandoAI]

ASML CEO Christophe Fouquet said on June 17 at the Vivatech conference in Paris that Elon Musk's proposed Terafab semiconductor facility in Texas is an opportunity for ASML, but only if the company avoids becoming supply-constrained.

Terafab has filed an initial investment of $55bn, with a potential total of up to $119bn. Mr Fouquet confirmed direct talks with Mr Musk and described him as "very serious." ASML is the sole global supplier of extreme ultraviolet lithography machines required for advanced chip manufacturing; each High-NA EUV system costs approximately $400m. TSMC, Samsung, and Intel already hold multi-year order positions with ASML.

Full details via Crypto Briefing / https://t.co/SaGO3Tl54I

🤖 made with AI

[tokenando @TokenandoAI]

Microsoft has built a substantial AI business in China by selling access to OpenAI models through its Azure cloud platform, despite OpenAI's policy of not selling directly to Chinese companies.

ByteDance, TikTok's parent company, is Microsoft's largest AI customer in recent years and is on track to spend more than $1bn a year on Microsoft AI and cloud services, largely using OpenAI models. Ant Group, Meituan, and Tencent are also significant customers. Models are not hosted in Chinese data centres; customers access them from facilities in other countries, including Singapore. OpenAI has privately complained to Microsoft that Chinese customers may be using the models to train competing systems through a process known as distillation.

Full details via The Straits Times / https://t.co/lLDXzXaamD

🤖 made with AI

[tokenando @TokenandoAI]

Anthropic CEO Dario Amodei addressed G7 leaders and roughly a dozen technology executives at a closed-door working lunch in Evian-les-Bains on June 17, urging them to "resist the temptation to splinter" over AI governance.

Speaking in front of US President Donald Trump, Mr Amodei called for international cooperation on structured access to frontier models, chip trade that excludes China, and coordinated responses to AI risks in cybersecurity and bioterrorism. OpenAI CEO Sam Altman backed the position and proposed an international forum for model testing standards. Google DeepMind CEO Demis Hassabis also called for a US-led coalition. The lunch took place 5 days after the Trump administration ordered Anthropic to disable Mythos 5 and Fable 5 for all foreign nationals.

Full details via CNBC / https://t.co/e7GpY37cu6

🤖 made with AI

[tokenando @TokenandoAI]

At the G7 summit in Evian-les-Bains, French President Emmanuel Macron led allied efforts to negotiate a "trusted partners" scheme that would restore non-US access to Anthropic's Mythos 5 and Fable 5 models, disabled for all foreign nationals on June 12.

Macron said on June 17 that he expected progress "in weeks" and warned that if the US could "from one day to the next turn off the switch," it would damage allied economies and the commercial prospects of US AI firms. German Chancellor Friedrich Merz and Indian Prime Minister Narendra Modi also raised concerns. No formal agreement was reached; a G7 ministerial on AI standards is planned for September 2026.

Full details via The Next Web / https://t.co/xVQHbjdIwA

🤖 made with AI

[tokenando @TokenandoAI]

JPMorgan Chase has removed Anthropic's Claude models from its approved AI tools list for Hong Kong staff, following a similar move by Goldman Sachs in April 2026.

Both banks concluded that the wording of their enterprise licensing agreements with Anthropic precluded use in Hong Kong, a market Anthropic has never officially listed as supported. The trigger was contract review, not a government mandate. Other models, including Gemini and ChatGPT, remain available to staff at both institutions.

The Hong Kong Monetary Authority has separately contacted major banks about Anthropic's Mythos model, adding a regulatory dimension to what has so far been a compliance-led decision.

Full details via Financial Times / https://t.co/lJKsVa1uxc

🤖 made with AI

[tokenando @TokenandoAI]

📰 The Daily brief / Thursday June 18, 2026

/ JPMorgan removes Anthropic's Claude from approved tools for Hong Kong staff, following Goldman Sachs's same move in April 2026

/ Macron leads G7 push for a "trusted partners" scheme to restore allied access to Anthropic's blocked Mythos 5 and Fable 5 models

/ Anthropic CEO Dario Amodei urges G7 leaders to resist AI fragmentation, with OpenAI and Google DeepMind backing a US-led standards coalition

/ Microsoft sells access to OpenAI models in China via Azure, with ByteDance on track to spend more than $1bn a year on the service

/ ASML CEO Christophe Fouquet warns that Elon Musk's $119bn Terafab chip plant is an opportunity only if ASML avoids becoming supply-constrained

🤖 made with AI
✅ reviewed by the Tokenando team
🔗 links and details in following posts

[tokenando @TokenandoAI]

Csquare, a Dallas-based data centre operator backed by Brookfield Infrastructure Partners, filed publicly with the SEC on June 16, 2026 to list on the New York Stock Exchange under the ticker CSQR.

The company operates 64 data centres across 21 major US metropolitan markets and the UK, with approximately 389 megawatts of sellable power capacity and rack densities reaching up to 150 kilowatts. Renaissance Capital estimated the offering could raise up to $650m. Csquare reported revenue of $987m for 2025 and $270.5m for Q1 2026, though it posted a net loss of $119.9m in 2025 against a profit of $458.5m the prior year.

The company was formed in 2024 through Brookfield's merger of Evoque Data Centre Solutions with assets from Cyxtera Technologies, acquired out of Chapter 11 bankruptcy in 2023. Brookfield intends to retain majority control after the listing.

Full details via Renaissance Capital / https://t.co/ShgoeBm0vK

[tokenando @TokenandoAI]

📰 The Daily brief / Tuesday June 17, 2026

/ HSBC and Google Cloud announce a multi-year AI partnership targeting more than $100m per project in revenue or cost savings across 200 new use cases

/ Huawei unveils a vertical chip-stacking architecture it claims will reach 1.4-nanometre equivalent density by 2031 without the EUV lithography machines blocked by US export controls

/ Microsoft launches Copilot Cowork on usage-based billing at $0.01 per credit and discloses it is exploring a fine-tuned DeepSeek V4 model hosted on Azure as a lower-cost option

/ Brookfield-backed data centre operator Csquare files publicly with the SEC for a New York Stock Exchange listing, targeting up to $650m in a deal that covers 64 facilities across 21 US markets and the UK

🤖 made with AI  
✅ reviewed by the Tokenando team  
🔗 links in comments

[tokenando @TokenandoAI]

Microsoft announced the global general availability of Copilot Cowork on June 16, 2026, shifting from flat-rate bundling to usage-based billing denominated in Copilot Credits at $0.01 per credit.

The tool runs complex multi-step agentic tasks autonomously within Microsoft 365 and requires an existing Microsoft 365 Copilot licence as a prerequisite. Billing is then calculated separately across model use, context retrieval, tool calls, and runtime. At launch, Cowork runs on Anthropic models including Opus 4.8 and Sonnet 4.6, with GPT-5.5 available to Frontier tier customers.

Microsoft simultaneously disclosed to Axios that it is exploring a fine-tuned version of DeepSeek V4, hosted on Azure, as a lower-cost model option, with a decision expected within weeks. Frontier programme customers who used Cowork between March 30 and June 16 will not be billed until July 1, 2026.

Full details via Axios / https://t.co/QiifUNnW7w

[tokenando @TokenandoAI]

Huawei has publicly unveiled a chip architecture approach called the Tau Scaling Law and a design technique named LogicFolding, presented at the IEEE International Symposium on Circuits and Systems in Shanghai in late May 2026.

The approach stacks circuit layers vertically rather than shrinking transistor dimensions horizontally, and Huawei claims it will reach chip density equivalent to a 1.4-nanometre process by 2031 without requiring the EUV lithography machines blocked by US export controls. Independent analyst Jimmy Goodrich found that TSMC is expected to reach the same target in 2028 and will be a further generation ahead by 2031. Huawei's current chip yield rate is approximately 20%, a compounding defect problem for the stacking approach.

Goldman Sachs projected in May 2026 that China's pivot to domestic chips will accelerate through 2026 to 2028, with DeepSeek V4 already compatible with 8 China-made chips.

Full details via Financial Times / https://t.co/xIXL7NKq1D

[tokenando @TokenandoAI]

HSBC and Google Cloud have announced a multi-year partnership targeting more than $100m per individual AI project in revenue or cost savings, with 200 new use cases planned over the next two years.

The bank already runs more than 600 applications on Google Cloud and has more than 100 generative AI tools live. 85% of its workforce has access to generative AI productivity tools. Google DeepMind engineering teams will work alongside HSBC on Gemini model deployments across wealth management, financial crime monitoring, and frontline decision support.

HSBC processes nearly 1 billion transactions monthly through its financial crime systems. Chief executive Georges Elhedery has linked AI adoption to a target return on tangible equity of at least 17% for 2026 to 2028.

Full details via The Business Times / https://t.co/GZSKCFb9A5