Latest edition · 19 November 2026Previous editions:5 to 12 November29 October to 5 November22 to 29 October15 to 22 October8 to 15 October1 to 8 October
Siam Kidd · Weekly briefing

Weekly Fascinating News Briefing

Covering 12 to 19 November 2026 · Macro, AI, agents, crypto, Bittensor, robots and new tech

Strongest conversation starters

Macro Economics

2 stories

A December rate rise now looks certain, and some at the Fed want a big one

Watch in 30 seconds

Minutes from the Fed's 27 to 28 October meeting, published on Wednesday, show that almost all officials think further rate rises are needed. A couple said a rise of more than 0.25 percentage points could be justified in December if inflation data stays bad. The Fed raised rates in September to a range of 3.75% to 4%, then held in October.

Fed Governor Christopher Waller told CNBC the same day that a rise of 0.5 points should be considered. "I'm very, very concerned about inflation," he said. "You might need to, like, shock people." The 10-year Treasury yield is back above 5%. The next decision is on 8 to 9 December.

What traders priced for 9 December on 17 November (CME FedWatch, per Investing.com)

Rise of 0.25 points (teal) vs 0.5 points (gold)97% and 3%

Chance of no change: about 0%.

Why it mattersHigher rates for longer are the main pressure on shares, bonds and bitcoin right now. A 0.5 point move would be the kind of shock Waller is openly talking about.

Jensen Huang is out at Nvidia, and sales still doubled

Watch in 30 seconds

Jensen Huang resigned as Nvidia's chief executive on 12 November after reports of several affairs, including one with a woman alleged to be a Chinese intelligence asset and one with a direct report. Finance chief Colette Kress is interim chief executive. Huang said on Tuesday he would be "honored" to return. Board chair Harvey Jones opened the results call by saying the board "does not intend to reinstate Jensen as CEO". Search firm Heidrick and Struggles is looking for a permanent replacement, and the board set up a new national security and export control committee.

Quarterly revenue was $115.8bn, up 103% on a year earlier, against about $109.2bn expected. Data centre sales were $107.6bn. Nvidia guided to $126bn to $134bn next quarter, assuming no data centre chip sales to China. Huang is due to testify to lawmakers on 1 December.

Quarterly revenue (to scale)

Expected (Bloomberg consensus)$109.2bn
Actual$115.8bn
Next quarter, top of guidance$134bn
Why it mattersNvidia is the centre of the AI trade. The numbers say demand has not slowed, but the company is now run by an interim boss while it investigates whether national security was put at risk.

AI

2 stories

Anthropic raised $81.3bn in the biggest stock market listing ever

Watch in 30 seconds

Anthropic priced its shares at $83 on Wednesday, above its own range of $70 to $78, and sold about 980 million of them. That raised $81.3bn (about £61.8bn), beating the $75bn SpaceX raised at first in June. Including employee shares and other dilution, the company is valued at about $1.5 trillion. Trading on Nasdaq under the ticker ANT starts today, 19 November.

Banks also have a 30-day option to buy up to 147 million more shares. Most of the growth behind the valuation came this year from Claude Code, its coding tool. OpenAI has filed confidentially to go public and says a listing is likely in 2027.

Money raised in the listing (to scale)

Anthropic, November 2026$81.3bn
SpaceX, June 2026 (before extra shares)$75bn
CaveatSpaceX ended up raising $85.7bn in total once its banks bought extra shares, so the record depends on whether Anthropic's banks use their option too.
Why it mattersThis is the first real public price for a frontier AI lab. How the shares trade in the first weeks will set the mood for OpenAI's listing and for AI shares in general.

92 public figures signed a statement saying "Anthropic is not a safe AI company"

Watch in 30 seconds

The Center for AI Safety published the "MisAnthropic Statement" on Wednesday, the day Anthropic priced its shares. The three-sentence statement accuses the company of building more and more autonomous AI while weakening its own safety promises and lobbying for rules that suit it.

The signers usually disagree with each other: safety campaigners like Max Tegmark and PauseAI, open-source figures like Hugging Face co-founder Thomas Wolf, and speed-up advocates like Guillaume Verdon and George Hotz, plus Steven Pinker, Judea Pearl and three Fields Medal winners.

CaveatNot every signer backs every word. Wolf said he would not have used the word "reckless" and did not know the statement's name when he signed.
Why it mattersAnthropic sells itself on being the careful lab. A mixed coalition attacking that claim on listing day gives investors a governance question to ask, not just a growth story.

New AI Models

2 stories

OpenAI's GPT-6.2 Astra does more work for less money

Watch in 30 seconds

OpenAI released two models on Wednesday: GPT-6.2 Astra for complex projects and GPT-6.2 Sol for everyday work. Both are in ChatGPT Work, Codex and the API. Astra's API prices are 20% lower than GPT-6 Astra's. On OpenAI's own GDPval v2 test of professional tasks, Astra scored 6.2 points higher while producing 22% fewer output tokens.

Astra is trained to produce work documents: slides that match a supplied template, spreadsheets, legal briefs and financial reports. On a coding test, OpenAI says the cheaper Sol finished within 0.5 points of Astra at about 20% of the cost.

CaveatThe benchmark and customer results are reported by OpenAI and its customers, not from a shared independent test.
Why it mattersThe race is moving from raw scores to cost per finished task. Cheaper, shorter answers matter more to businesses than another benchmark record.

Google's Gemini 4 Hydrogen is a cheap, fast model for coding agents

Watch in 30 seconds

Google DeepMind launched Gemini 4 Hydrogen on 13 November, priced at $1.50 per million input tokens and $7.50 per million output tokens, with a context window of about 1 million tokens. Google says it finishes tasks 2.5 times faster than Gemini 3.8 Flash and makes more than five times fewer made-up claims.

It scores 51 on the Artificial Analysis Intelligence Index, just ahead of Gemini 4 Argon. It is now the default model for Google's Antigravity coding agent, which can also hand planning to a bigger model and the building work to Hydrogen.

Terminal-Bench 4.1 score (to scale)

Gemini 4 Hydrogen57%
Gemini 3.8 Flash18.2%
CaveatClaude Sonnet 5.5 still scored higher on Terminal-Bench (63.9%) in Google's own comparison. "Tokens" are chunks of text, each roughly 75% of an average word.
Why it mattersCheap models that are good enough for most coding work are what make always-on agents affordable.

AI Agents: Capability, Reliability and Safety

2 stories

ChatGPT's new healthcare agents click through hospital and insurer systems

Watch in 30 seconds

OpenAI added computer-use agents to ChatGPT for Healthcare on 16 November. They read screens, click and type, so they can move information between electronic health records and insurer websites even where the software has no proper connection. It also added more than 20 healthcare plugins and better clinical search.

In a demo with made-up patient data, a clinician asks by voice for help with a back-pain MRI. The agent reads the notes and past scans, drafts the order, a visit note and a reply to the patient, then prepares the insurer's approval request for a person to check. CVS Health says one agent cleared a months-long backlog of documents over a weekend, and Cigna is using it too.

CaveatThe demo used made-up data. OpenAI says doctors rated 99% of 325 reviewed tasks safe, but those are its own tests, and a person still checks each step.
Why it mattersPaperwork, not diagnosis, is where most healthcare time goes. Agents that can work the screens staff already use could cut that without waiting for software makers to connect their systems.

Perplexity now sells one key for every major AI model, with automatic backup

Watch in 30 seconds

Perplexity launched its Orchestrator API on 18 November. Developers get models from OpenAI, Anthropic, Google and xAI, plus open models like GLM, Kimi and DeepSeek, through one key and one bill. If one provider errors, hits a limit or times out, requests switch to another copy of the same model automatically.

It accepts the standard OpenAI and Anthropic request formats, so existing apps can switch over with small changes. Perplexity says the system behind it already routes hundreds of millions of requests across more than 85 models and over a dozen providers.

Why it mattersAgents that run all day break when one provider has an outage. A single switchboard with failover makes them more reliable, and makes Perplexity the middleman.

Crypto

3 stories

Bitcoin fell to a four-month low near $62,500

Watch in 30 seconds

Bitcoin fell 3.5% on Wednesday to about $62,486, its lowest since 13 July, after the Fed minutes. It fell more than 15% the week before, and is down from a peak of about $86,000 in early October. Ether dropped to about $1,770.

Market maker Wintermute warned on Tuesday that prices could slide into the $50,000s, with $62,000 the first level to watch and $60,000 the next. It blames rate rise bets and a 10-year yield above 5%, plus money leaving bitcoin funds and stablecoins. It would turn less cautious if bitcoin closed above $68,000 with fund inflows.

Bitcoin price levels (to scale)

Early October peakabout $86,000
200-week averageabout $68,000
18 November low$62,486
Why it mattersBitcoin is behaving like the riskiest bet on interest rates. Until the December rate decision is clear, the pressure is mostly macro, not crypto news.

Strategy kept buying as its paper loss hit $9.6bn

Watch in 30 seconds

Strategy bought another 485 bitcoin for $33.4m last week at an average of $68,887 each, taking it to 850,849 bitcoin bought at an average of $75,452. With bitcoin near $64,000, The Block puts its paper loss at $9.6bn. It also spent $228.1m buying back its STRC preferred shares.

MSCI will remove Strategy from its indexes, including the MSCI USA Index, after the close on 30 November, under a new rule for companies that look more like investment funds. Strategy estimates funds tracking those indexes hold about 13 million of its shares, 3.1% of the total.

Why it mattersForced selling by index funds could shrink the premium Strategy's shares trade at, which is what lets it raise money to buy more bitcoin.

Democrats plan a tougher crypto rulebook after winning Congress

Watch in 30 seconds

After Democrats won both houses in the November midterms, House Democrats are discussing their own crypto market rules for the next Congress, based on Maxine Waters' Crypto Investor and Consumer Bill of Rights. Ideas include a bigger role for the SEC and a smaller one for the CFTC, which is the opposite of the industry-backed CLARITY Act. Ethics limits on crypto dealings by the President and other officials are also likely.

The CLARITY Act passed the House in 2025 but failed a Senate vote in September, and it lapses when this Congress ends on 3 January 2027. Republicans call the Democratic plan a "messaging bill".

CaveatNo bill text exists yet. The details come from industry sources speaking to journalist Eleanor Terrett, not from lawmakers.
Why it mattersClear US crypto rules now look years away, and the next version could be stricter on exchanges and token issuers.

Bittensor

3 stories

Quantum-resistant wallets are now switched on, and custody gets new safeguards

Watch in 30 seconds

Last week's story was that post-quantum wallets were on the chain but not switched on. That has changed: Taostats' protocol note, last reviewed on 17 November, says root has turned the feature on and the new accounts can register and sign transactions. The TAOApp wallet extension now supports the new key types. One community member suggested ML-DSA-65 for most people, or the cheaper Falcon-512 for frequent traders.

Two more releases followed after the SN3 (Teutonic) owner's coldkey was compromised on 11 November. Release 491 lets any holder set up a protected account with delayed authority over funds and stake. Release v492 changes how funds move during a coldkey swap and stops a swap overwriting an address that already holds funds.

CaveatHardware wallets, Polkadot Vault and other tools need their own support. Check your wallet software before moving any TAO.
Investor takeawayBittensor now has quantum-safe accounts in use and better tools against stolen keys. Holders still have to choose to move to them.

Lium says it bought back and burned $3.3m in a week from real revenue

Watch in 30 seconds

SN51 (Lium), a GPU rental subnet, reported about $3.3m of buybacks and burns over the past week, all paid for by customer revenue, with about 500 Nvidia B300 GPUs almost fully rented. It also logged a record day of about $129,000 in customer spending.

A week earlier the team reported a 30-day jump in paying customers of 78% and in deposits of 94%, a buy-and-burn of about 5,000 TAO, and a cut in the owner's share of emissions from 18% to 9%.

Lium's reported numbers

$3.3m
Buybacks and burns in a week
~500
B300 GPUs, almost fully rented
18% to 9%
Owner share of emissions
CaveatThese figures are reported by the Lium team and have not been independently audited.
Investor takeawayThe community's big complaint this week was that much subnet "revenue" is emissions going round in circles. Lium is the clearest claim of outside money flowing back into a token.

Chutes signs Tencent Cloud and Alibaba Cloud in one week

Watch in 30 seconds

SN64 (Chutes) announced a partnership with Alibaba Cloud on 18 November, days after one with Tencent Cloud. The Alibaba deal brings Qwen3.8 Max, GLM-5.3 and DeepSeek V4.1 Flash to developers through Chutes, with 35% off the last two until 23 November.

The same day, Taostats launched live cross-chain swaps, letting people swap USDC or ETH for TAO or subnet tokens without a bridge or an exchange account.

Investor takeawayTwo of Asia's biggest cloud firms now sell through a Bittensor subnet, and buying TAO just got easier for anyone holding stablecoins.

Robots

3 stories

Nvidia released free software that turns human movement into robot moves

Watch in 30 seconds

Nvidia released SOMA Retargeter, an open-source toolkit that adapts recorded human movement to a robot's different limb lengths and joints. Version 0.2 comes ready for five humanoids: Unitree G1 and H2, Booster T1, AgiBot X2 Ultra and AgiBot A3T3.

It outputs the robot's position and joint angles for every frame, which a control policy then learns to follow in simulation before trying it on real hardware.

Why it mattersFilming people is far cheaper than programming robots by hand. Free tools like this lower the cost of teaching humanoids new skills.

Google's robot plugin takes a new arm from box to working app in an afternoon

Watch in 30 seconds

Google launched a Robotics plugin for its Antigravity coding tool on 18 November. Paired with Gemini 4 Hydrogen, developers describe what they want in plain English, and it installs drivers, calibrates the arm and camera, builds a simulation and writes the control code.

In Google's demo, an arm and a depth camera were calibrated and ready in about 10 minutes after being plugged into a laptop. The same code is tested in the MuJoCo simulator, then run on the real arm.

CaveatThese are Google's own demos on supported arms, not independent tests.
Why it mattersSetting up a robot used to take specialist engineers days. If plain-language setup works, far more small firms can try automation.

Boston Dynamics' Atlas humanoid is heading into chip factories

Watch in 30 seconds

GlobalFoundries will become the first company outside the car industry to test Boston Dynamics' Atlas humanoid, with on-site testing starting in early 2027. It is also rolling out Spot robot dogs for inspections in New York, Vermont, Dresden and Singapore, where Spot has done more than 90,000 inspections since April.

Atlas can lift 30kg continuously and 50kg briefly, reach 2.3 metres, and swap its own battery.

Atlas and Spot in numbers

30kg
Continuous lift
2.3m
Reach
90,000+
Spot inspections in Singapore since April
Why it mattersHumanoids are moving from car plant pilots into other high-value factories, where they can work alongside existing equipment.

New Technology

3 stories

Nvidia says a new algorithm speeds up quantum error-correction research 13,000 times

Watch in 30 seconds

Nvidia released a decoding algorithm called TCBP-JOBS on 16 November and says it runs up to 13,000 times faster than leading methods, cutting some quantum computer design research from 12 weeks to 45 minutes.

It also announced a quantum design platform with IBM and the US Fermilab lab, linking Nvidia DGX B200 systems to IBM quantum computers.

Nvidia's claim in numbers

13,000x
Faster decoding, at best
12 weeks
Research run before
45 min
Same run now
CaveatThe speed-up is Nvidia's own figure, and it applies to research simulations, not to running quantum computers.
Why it mattersError correction is the main thing stopping quantum computers being useful. Faster research tools shorten that path.

UCL found a way to pack more energy into safer water-based batteries

Watch in 30 seconds

A UCL-led team, writing in Nature on 13 November, made iodine in a water-based battery move four electrons instead of two each time it charges, which raises how much energy it can store. A new electrolyte keeps water away from the iodine to stop damaging side reactions.

In zinc-iodine batteries it kept almost all its capacity after thousands of charge cycles, and worked in pouch cells like those in phones and laptops.

Why it mattersWater-based batteries cannot catch fire like lithium-ion ones. More capacity makes them a more realistic choice for grid and home storage.

A diamond maser boosts quantum signals over 1,000 times with almost no noise

Watch in 30 seconds

Researchers at the Okinawa Institute of Science and Technology built a diamond microwave amplifier that works near absolute zero and boosts signals more than 1,000 times. It adds only 2.55 units of noise, close to the physical minimum of 0.5.

Unlike the amplifiers used in most quantum computers today, it handles much stronger signals and works inside magnetic fields. It is now being developed towards a product.

Why it mattersReading out qubits cleanly is a bottleneck for quantum computers and sensors. A tougher amplifier could make both more practical.