█▓▒░
█████▓▒░
██████████▓▒░
███████████████▓▒░
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. 21 · Thu · 9 July 2026
Happy Thursday. OpenAI's most capable models go on public sale today,
two weeks after Washington asked the company to keep them inside a
narrow circle of approved organizations. Elon Musk's SpaceXAI answered
on Wednesday with a model priced at a fraction of Anthropic's flagship.
Beijing, meanwhile, is quietly reopening the door to Nvidia chips, at
the very moment Nvidia itself has shed a trillion dollars in market
value.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░
OpenAI puts GPT-5.6 on public sale today after Washington
lifts its hold (update)
All three GPT-5.6 models go on public sale globally today: Sol, the
flagship, at $5 per million input tokens and $30 per million output;
Terra, which OpenAI says performs like GPT-5.5 at half the cost; and
Luna, the cheap one. When the family launched in late June we covered it
going to a government-cleared few: OpenAI had paused the wide rollout at
the White House's request, in consultation with the Office of Science
and Technology Policy and the Office of the National Cyber Director,
holding early access to about 20 vetted organizations. Axios first
reported the clearance, and says the testing was carried out by the
Center for AI Standards and Innovation, a unit inside the Commerce
Department, with OpenAI keeping technical staff in Washington through
the review. Covered by Axios, CNBC and PYMNTS (PYMNTS
· Axios · CNBC).
SpaceXAI and Cursor ship Grok 4.5, priced at a quarter of
Anthropic's Opus
Grok 4.5 landed Wednesday, the first model jointly built by Musk's
SpaceXAI and Cursor, the coding startup SpaceX is buying for $60 billion
in stock. Cursor calls it "our most intelligent model and the first
we've built for more than software engineering," and it was trained on
trillions of tokens of Cursor users' own code. At $2 and $6 per million
input and output tokens, it undercuts Anthropic's Opus 4.8 about
fourfold on output and OpenAI's new Sol fivefold; Musk describes it as
Opus-class. Covered by Bloomberg and TechCrunch (Cursor).
Beijing will let its top AI labs buy Nvidia H200s again, for
training only
Chinese officials have told Alibaba, ByteDance and DeepSeek in recent
weeks that they may soon be allowed to buy Nvidia's H200 accelerators,
The Information reported Wednesday, after a year in which those firms
made do with smuggled chips, cut-down export models and Huawei silicon.
The quantities are rationed rather than generous: Beijing is weighing
approval for fewer than 200,000 chips in total, less than half what the
companies asked for earlier this year, with inference workloads still
expected to run on domestic processors. The loosening is narrow, and it
comes as Beijing weighs new limits on the models its own labs would
open-source, the export clampdown we covered yesterday (more in AI,
below) (The
Information · Reuters).
Nvidia has lost $1 trillion in market value in under two
months
Nvidia is down 16% from the all-time high it set on May 14, a slide
that leaves the stock at 18 times projected earnings, its cheapest since
early 2019, before the AI boom began. Analysts have not cut their profit
estimates and the company still holds roughly 97% of the server GPU
market. What has changed is where investors want exposure: Micron is up
229% this year on memory demand, while Nvidia is up 5.6%, trailing both
the S&P 500 and the Nasdaq 100. The firm selling the picks is being
outrun by the firm selling the memory (Bloomberg
via Yahoo Finance).
Blue Origin takes $10 billion of outside money, 26 years
after Bezos started funding it alone
Jeff Bezos is raising $10 billion for his rocket company at a $130
billion pre-money valuation, the first external round since he founded
it in 2000. Coatue is putting in about $4 billion and Bezos another $2
billion of his own. The raise follows the explosion of a New Glenn
rocket during testing in late May, and lands in a market that has just
watched SpaceX list at $1.75 trillion. Reported by the New York Times
and TechCrunch (TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░
- $1 trillion — Nvidia market value erased in under
two months (Bloomberg)
- 200,000 — H200 chips Beijing may approve, under
half what its labs requested (The Information)
- $130 billion — pre-money valuation in Blue Origin's
first outside funding round (TechCrunch)
- $6 — Grok 4.5's price per million output tokens,
against $30 for GPT-5.6 Sol (Cursor · OpenAI)
- 8,000 — Discord accounts banned since May by a
broken AI moderation filter (TechCrunch)
░░▒▒▓▓ AI ▓▓▒▒░░
- Anthropic gave Fable 5 a stay of execution on Tuesday (update),
pushing back the July 7 cutoff we reported last week: the model stays
bundled into Pro, Max, Team and premium Enterprise seats until July 12,
after users objected. Subscribers can spend half their weekly allowance
on it; after Sunday they buy usage credits or drop to a cheaper model
(Android
Authority).
- Shanghai's MiniMax is building a 2.7-trillion-parameter model,
internally called M3 Pro, roughly six times the size of the M3 it ships
today and larger than any Chinese model on the market. It plans to
open-source it, possibly in the third quarter, which is precisely the
sort of release Beijing's commerce officials have been discussing
restricting (The
Decoder).
- OpenAI released GPT-Live on Wednesday, voice models that listen and
speak at the same time, instead of waiting for you to finish. The
full-duplex design is what makes live translation possible, and the
smaller version replaces Advanced Voice Mode as the default in ChatGPT
(TechCrunch).
░░▒▒▓▓ TECH ▓▓▒▒░░
- Anyone can now tag a public Instagram account in a Meta AI prompt
and pull that person's photos into a generated image. Public accounts
are opted in by default, the switch to turn it off sits four menus deep
under Sharing and reuse, and flipping it stops only future generations.
Meta's own help page, as WIRED notes, puts it plainly: "You will not be
notified about content created using AI features at Meta." (WIRED).
- SambaNova, which builds inference chips meant to run
trillion-parameter models on a single rack, closed the first $1 billion
of a Series F at an $11 billion valuation, led by General Atlantic.
JPMorgan Chase signed on as an inference customer and will run the
systems on its own premises instead of renting them from a cloud (TechCrunch).
- Discord said on Tuesday that a bug in its AI moderation system had
wrongly banned more than 8,000 accounts since May, after the filter
matched harmless uploads against a database of known abuse material.
Among the images that got people thrown off the platform: spreadsheets,
chessboards, game textures and blank gray backgrounds. Everyone is being
reinstated (TechCrunch).
░░▒▒▓▓ POLICY ▓▓▒▒░░
- The EU's General Court threw out Apple's challenge to its
designation as a gatekeeper under the Digital Markets Act, keeping iOS
and the App Store bound by rules on rival app marketplaces, developer
payments and interoperability. Apple had argued its five app stores were
five separate services; the judges found "those stores have the same
purpose, namely to connect app developers with end users." An appeal on
points of law remains open (Euronews).
- The AI image generator Midjourney is trying to turn Hollywood's own
AI habits into a defense. Sued by Disney, Universal and Warner Bros.
over copyrighted characters, it has asked the court to make the studios
hand over every prompt their staff typed into Midjourney and the
internal AI work behind their storyboards. A magistrate limited
disclosure to consumer-facing products in June; Midjourney wants that
reversed, because if the studios train and prompt as it does, that
starts to look like industry custom, which is one of the things courts
weigh in a fair-use fight (Techweez).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░
- Mercor, which pays domain experts to write the training data
frontier labs cannot scrape, passed $2 billion in gross annualized
revenue in June, four months after its first billion. Roughly two-thirds
of that goes straight back out to contractors, and the company says it
is now free-cash-flow positive (The
Information).
- Amazon sold at least $25 billion of bonds on Tuesday across eight
tranches maturing between three and 40 years, to help fund roughly $200
billion of capital spending this year against $131 billion in 2025.
Orders peaked near $62 billion, then fell to about $41 billion once the
banks trimmed the extra yield on offer. Amazon has told its underwriters
it will not borrow again in 2026 (SiliconANGLE).
- Chip-and-AI research firm SemiAnalysis published its own
reconstruction of Anthropic's finances on Wednesday, five weeks after
the company filed confidentially for an IPO. On SemiAnalysis's model,
Anthropic clears $1 billion of operating profit in the current quarter
and could, if it keeps executing, become the first $6 trillion company.
The numbers are estimates built from the outside, not disclosures (SemiAnalysis).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░
- President Trump declared the interim agreement with Iran finished on
Wednesday (update), hours after the US strikes we reported yesterday
answered attacks on three ships in the Strait of Hormuz. "For me, I
think it's over," he said, while allowing that talks could continue.
Brent crude rose more than 5% to above $78 a barrel (ABC7).
- At the NATO summit in Ankara, which we previewed on Monday, Trump
called for an end to all US trade with Spain, the one member that has
not committed to the alliance's 5%-of-GDP defense target and that
refused American forces the use of its bases and airspace during the
Iran war. He also renewed his demand for control of Greenland, prompting
Danish Prime Minister Mette Frederiksen to reply that "Greenland is, of
course, not for sale." (ABC
News).
- Marine Le Pen announced on Tuesday evening that she will run for the
French presidency in 2027 (update), hours after the appeals court ruling
we covered yesterday shortened her ban from public office. The court
attached an electronic ankle tag, a condition she had said would rule
her out; she is appealing to France's highest court and says she need
not wear it while that appeal is pending (France
24).
░░▒▒▓▓ ALPS ▓▓▒▒░░
- Switzerland does not want much international AI regulation,
communications minister Albert Rösti told reporters in Geneva on
Wednesday, days after the UN's first Global Dialogue on AI governance
met there, as we covered yesterday. Governance is welcome so long as it
stays minimal, he said: the UN should set a floor, and the industry
should certify itself against best practice. Bern has already declined
to write an AI act of the European kind (SWI
swissinfo.ch).
- Munich's QuantumDiamonds raised €15 million led by climate fund
World Fund, and a further €76 million in non-dilutive money from
Germany's federal economy ministry and the state of Bavaria under the
European Chips Act. Its synthetic diamonds sense magnetic fields well
enough to find defects through every layer of a stacked chip without
destroying it, in about two minutes instead of weeks (TechCrunch).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░
- A SpaceX rideshare carried the first commercially built
nuclear-powered satellite into orbit on Tuesday. BOHR, a cubesat from
Florida's City Labs, runs its bus on ordinary solar cells; the payload
it flew to prove is powered by the beta decay of tritium, and keeps
working when the sun goes down (Space.com).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░
- Mistral released Robostral Navigate, its first model for the
physical world: an 8-billion-parameter system that takes a
plain-language instruction and a feed from one ordinary camera, and
drives a robot to the right place. Trained entirely in simulation on
400,000 trajectories, it scores 76.6% on the standard R2R-CE navigation
test, beating the best single-camera rival by 9.7 points and, more
awkwardly for the sensor industry, the best lidar-and-depth system by
4.5. It runs on wheeled, legged and flying robots alike (Mistral).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░
- HubSpot told customers on July 1 that it would begin pooling their
CRM data, including business contacts and employer details, to power a
lead-finding feature due in August, with the setting on by default and
three separate switches to turn it off. Customers revolted on LinkedIn;
on July 5 chief product and technology officer Duncan Lennox scrapped
the change, writing "We made a mistake." Takeaway: a default is a
statement of values, and Lennox's apology landed because it named the
decision, not the backlash (CX
Today).
- A Seattle startup called Optimly has raised $800,000 to run a public
index where brands can claim and correct what chatbots say about them,
on the premise that AI models learn about a company from Reddit,
Wikipedia and reviews, not from its own website. More than 100 brands
have claimed profiles, the index now fields around 11,000 agent requests
a week, and Anthropic's Claude has started reading from it. Takeaway:
the newest corporate-communications channel has no journalists in it at
all (GeekWire).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░
Two teenagers, 14 and 15, took a Waymo through San Mateo on Monday
afternoon, drinking and firing an Orbeez pellet gun out of the window.
Staff watching the interior cameras saw what they took for a recoil,
called 911, and stopped the car near El Camino Real. To keep the
passengers inside until officers arrived, the company told them the
vehicle was "experiencing mechanical trouble." The district attorney is
deciding whether to charge them (NBC
Bay Area).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░
🎵 Flicker —
Batu, Donato Dozzy
Buckle up — from the Uncertain Futures desk