░▒▓██▓▒░
░▒▓████████▓▒░
░▒▓████████████████▓▒░
░▒▓████████████████████████▓▒░
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
════════════════════════════════════════
No. 20 · Wed · 8 July 2026
Happy Wednesday. Nvidia's next AI server rack may not arrive until
2028, chip researchers say, though the company insists otherwise, even
as American companies quietly route a third of their AI work to cheaper
Chinese models. Anthropic, meanwhile, says it can now see what Claude is
thinking before it speaks. In Geneva, the UN opened its first summit on
governing all of this, where the Secretary-General called autonomous
killer robots morally repugnant.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░
Nvidia pushes its next AI server rack to 2028, and chip
suppliers slide
Nvidia's Kyber rack, the system meant to wire 144 of its 2027 Rubin
Ultra chips into a single machine, has slipped more than a year to 2028,
chip-research firm SemiAnalysis reported Monday, undone by a 78-layer
circuit board that is proving nearly impossible to manufacture. The
Rubin Ultra, meant to fuse four silicon dies into a single processor,
was cut to a two-die design, and a stopgap rack was scrapped after cloud
customers balked, leaving Nvidia, on SemiAnalysis's read, without a
proven way to scale up that generation and handing AMD and Google a rare
opening. Nvidia countered that its roadmap remains intact; unconvinced
Asian circuit-board and memory suppliers sold off. Covered by CNBC and
SemiAnalysis. (CNBC)
US companies now route a third of their AI to Chinese models
— and Beijing may pull the export plug
The share of tokens US firms run through Chinese open models on the
OpenRouter marketplace has topped 30% every week since February, up from
an 11% average over the prior year, as DeepSeek, Z.ai and Alibaba
undercut OpenAI and Anthropic by 60% to 90% on price. Now the traffic
may reverse: Chinese officials led by the Ministry of Commerce are
weighing curbs on overseas access to the country's most capable models,
open and closed alike, mirroring the export controls Washington placed
on Anthropic's cyber-capable Mythos. Wharton's Ethan Mollick, posting on
Bluesky, doubted the flow of frontier open-weight models would last much
longer. Covered by CNBC and Reuters. (CNBC
· Reuters)
The UN opens its first AI-governance summit in Geneva, and
Guterres calls for a ban on killer robots
All 193 member states sat down together in Geneva on Monday for the
UN's first Global Dialogue on AI Governance, and Secretary-General
António Guterres used the opening to draw red lines. He called lethal
autonomous weapons that pick and kill targets without human judgment
"morally repugnant," pressed for common safety standards, and warned
that the world cannot vibe-code the future of humanity. The catch: the
dialogue can set norms, but it binds no one. Covered by the WSJ and UN
News. (UN
· WSJ)
Anthropic says it can now read the thoughts Claude doesn't
say out loud
In interpretability research out Monday, Anthropic says it found a
small internal workspace it calls J-space that "operates silently, in
the model's internal neural activations," holding the concepts Claude
reasons through before any reach the answer. Swap the model's hidden
concept from spider to ant mid-thought and its leg count flips from
eight to six; in a model secretly trained to sabotage code, words like
fake and fraud surface in the workspace while the visible output stays
clean. On Anthropic's own telling the work says nothing about whether
Claude is conscious, but it hands safety researchers a way to catch a
model scheming before it speaks. Covered by Axios and The Rundown. (Anthropic)
Four US states want $1.4 trillion from Meta — roughly its
entire market value
Meta said in a court filing that California, Colorado, Kentucky and
New Jersey are seeking about $1.4 trillion in penalties at an August
youth-safety trial, a sum close to the company's whole market
capitalization. The states allege Meta engineered Facebook and Instagram
to hook minors and hid the risks; Meta calls the figure untethered from
any real harm. Whatever a jury makes of a number that large, it resets
the ceiling on what platform-design liability might cost. Reported by
Reuters. (Reuters)
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░
- more than 30% — the share of US companies' AI token
use now running on Chinese models (CNBC)
- $1.4 trillion — penalties four US states are
seeking from Meta at an August trial (Reuters)
- 63% — solo founders' share of new C-corps formed
through Stripe last quarter, an all-time high (Stripe)
- $2.5 billion — the valuation at which Agility
Robotics becomes the first US pure-play humanoid maker to go public (The
Robot Report)
- €411 million — Proxima Fusion's raise, Europe's
largest fusion round on record (Bloomberg)
░░▒▒▓▓ AI ▓▓▒▒░░
- Microsoft joined the AI cost-cutting turn, starting to swap its own
models in for OpenAI's and Anthropic's on high-volume tasks inside
products like Excel and Outlook, while an internal memo reset Copilot
around a blunt standard (it must "earn the right to exist") and folded
its consumer and enterprise apps into one superapp (The
Information).
- Meta shipped Muse Image and previewed Muse Video, the first image
and video models from its Superintelligence Labs, dropping them free
into Instagram, WhatsApp and Meta AI, its bid to catch OpenAI and Google
in consumer image generation (CNBC).
- Two Chinese labs went a step past buying silicon to designing it,
the in-house route Anthropic and OpenAI are already on: DeepSeek, which
just last week was adapting its models to run on Huawei chips, is now
designing its own inference chip to depend less on both Huawei and
Nvidia, and Zhipu is weighing a custom chip of its own as demand for its
GLM models strains its compute (The
Information · Reuters).
░░▒▒▓▓ TECH ▓▓▒▒░░
- Nvidia's answer to the startups gunning for it is to sign them up:
it is combining its hardware with inference-chip challenger d-Matrix in
a joint system, the latest of several deals that turn would-be rivals
into partners (The
Information).
- Elon Musk renamed xAI SpaceXAI this week, five months after folding
it into SpaceX, tucking the Grok maker into the rocket company's story
at a valuation the merger set near $250 billion (Gizmodo).
- The US cyber agency CISA is running Anthropic's Mythos model over
federal code to hunt security holes, and finding plenty, Reuters
reported this week, even as researchers documented JadePuffer, billed as
the first ransomware attack driven by an AI agent, which on closer
inspection still needed a human in the loop (Reuters
· TechCrunch).
░░▒▒▓▓ POLICY ▓▓▒▒░░
- The ITU AI for Good Global Commission we flagged last week is
convening in Geneva this week (update): the 40-plus-member body,
co-chaired by Salesforce's Marc Benioff and Rwanda's Paul Kagame with
tech chiefs like Jensen Huang and Andy Jassy, opens its first session
alongside the states-only Global Dialogue up top, seating many of the
people building AI at the table meant to govern it. No commitments have
surfaced yet (ITU).
- Australia's world-first under-16 social-media ban is failing at the
first step (update): testers opened 50 accounts claiming to be 16 and
nine of ten platforms asked none of them for proof of age, a study seen
by Reuters found; the government had already doubled the maximum fine
last month (NBC
News).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░
- The skeptics have company at the top: US Treasury analysts drafted
an internal report warning that an AI bubble could ripple through
data-center financing, private credit, chipmakers and utilities, and
separately a Nobel-winning economist told Bloomberg that AI will not
restore the West's lost era of rapid productivity growth (NOTUS
· Bloomberg).
- A federal judge on Monday refused to throw out the jury verdict that
found Elon Musk defrauded Twitter investors by muddying its bot numbers
before the 2022 takeover, leaving him exposed to about $2.6 billion in
damages (Al
Jazeera).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░
- US forces struck more than 80 targets in Iran after Iranian
projectiles hit three tankers in the Strait of Hormuz, and the Treasury
tore up the oil-sanctions waiver it granted last month (update); the
frayed US–Iran truce we have tracked finally snapped as Iran prepared to
bury Supreme Leader Ali Khamenei (CBS
News).
- Ukraine flew its deepest strike of the war on Monday, hitting the
Omsk refinery, Russia's largest, some 2,500 kilometers away and worth
roughly a tenth of national refining capacity, while a shortage of
Patriot interceptors let Russia's ballistic missiles through over Kyiv
(update) (CNBC).
- A Paris appeals court shortened Marine Le Pen's ban from public
office, clearing the far-right leader to run for president in 2027.
Because a further appeal suspends her sentence, she would campaign
without the electronic ankle tag it otherwise carries (CNN).
░░▒▒▓▓ ALPS ▓▓▒▒░░
- Munich's Proxima Fusion raised €411 million ($468 million) at a €2.4
billion valuation, Europe's largest fusion round on record, with Google
making its first European fusion bet alongside utility RWE, to build a
stellarator, a twisting magnetic reactor, in Bavaria (Bloomberg).
- WEF founder Klaus Schwab told authorities he found a hidden
listening device in his Geneva home this week and filed a criminal
complaint, the latest turn in a bitter fight over the forum's future, a
year after he left amid a misconduct probe. The man whose gathering runs
on discretion is, he says, the one now being bugged (Bloomberg).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░
- A French lab and its partners taught a neural network to be a video
game: Kyutai, General Intuition and Epic Games released MIRA, an
open-source world model (a system that generates an environment frame by
frame), running playable 2v2 Rocket League for four players with no game
engine underneath, learned from 10,000 hours of bot matches. Its memory
lasts about four seconds, so when a goal replay rolls, it cheerfully
invents a goal that never happened (MIRA).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░
- Agility Robotics agreed to go public by merging with the blank-check
shell Churchill Capital Corp XI at about a $2.5 billion valuation,
making the Oregon maker of the Digit warehouse robot the first US
pure-play humanoid company to trade publicly and raising more than $620
million, with Foxconn leading a $200 million placement (The
Robot Report).
- Forterra has more than 100 self-driving Lancer ATVs running supply
and casualty-evacuation missions in Ukraine, the first American
autonomous ground vehicles in combat, though soldiers mostly teleoperate
them, the machines being too valuable to lose and full autonomy not
quite ready for war (TechCrunch).
░░▒▒▓▓ REPORTS ▓▓▒▒░░
- Future of Life Institute — AI Safety Index (Summer
2026). The safety group's expert panel graded the frontier labs
and handed out no A's: Anthropic led with a C+, OpenAI and Google
DeepMind trailed with C's, and Musk's SpaceXAI drew an F, with reviewers
faulting the labs for quietly dropping earlier pledges to pause if their
systems grew too dangerous. "AI companies are sprinting toward a cliff,"
said FLI chair Max Tegmark (Future of
Life Institute).
- Stripe Economics — The Age of the Solopreneur.
Stripe's data team makes the case that AI is minting one-person
companies: solo founders were 63% of the C-corps formed through Stripe
Atlas last quarter, an all-time high, and more than twice as many
solopreneurs cleared $1 million in revenue in 2025 as in 2023 (Stripe
Economics).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░
- Reddit is fighting AI with AI, deploying its own models to scrub a
wave of stealth-marketing posts, brand-planted content designed to get a
product name in front of ChatGPT and Gemini, which increasingly learn
from Reddit threads, Bloomberg reported this week. Takeaway for
communicators: as answer engines mine community forums, gaming them is
becoming its own discipline, and the platforms are now building
detectors faster than the spammers can plant (Bloomberg).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░
Norway's Erling Haaland has been all over the World Cup this summer.
Except a lot of it isn't him. WIRED reports the striker has become an
internet character, his likeness spun into an endless stream of
AI-generated clips and memes by fans faster than anyone can keep track.
Somewhere between the real goals and the synthetic ones, the tournament
has acquired a second Haaland who never laced a boot (WIRED).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░
🎵 (̸̢̛̼̞̭͋ͅ)̸͚̰͛̔̾̀̿͒͂v̴̢͚͚͎ȯ̶̞̮͖̑̈́OOOOOOOooo☼⃝☼⃝◉࿃ूੂ࿃ूੂੂ
ʅ͡͡͡͡͡͡͡͡͡͡͡()ʃ͡͡͡͡͡͡͡͡͡͡ ꐑ(ఠీੂȯ̶̞̮͖̑̈́̿)̸̳̥̰̜̥̺̐ͅ࿃ूੂ✧⃛✧⃛)̴͎̜͍̱̋̌͋̓̾̚͜ ̷̨̢̥̅͝ͅ(̸̢̛̼̞̭͋ͅ)̸͚̰͛̔̾̀̿͒͂:̴͓̞̑̌̂̆̊͋̀:̸͎̟̯̂̓̌:̶̢͙͙͕̠̩͆(̷̮͍͚̫͚͂̍̒̋̕)̵̳̗̊̍̽́ ̵͍̪̲̀̍ ̴̤̚ ̴̹̙͙̻̓̉̚͝ ̟̞̝̜̙̘̗̖҉̵̴̨̧̢̡̼̻̺̹̳̲̱̰̯̮̭̬̫̪̩ — ⣎⡇ꉺლ༽இ•̛)ྀ◞ ༎ຶ ༽ৣৢ؞ৢ؞ؖ ꉺლ
Think freely — from the Uncertain Futures desk