AI & Software
Where AI spending is actually landing
Model releases, enterprise adoption and software margins, read for what they do to revenue — across the labs, the platforms and the software they run on.
2026-10-23
%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%22220%22%20cy%3D%22150%22%20r%3D%22180%22%20fill%3D%22%2322d3ee%22%20fill-opacity%3D%220.12%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%221010%22%20cy%3D%22130%22%20r%3D%22220%22%20fill%3D%22%2338bdf8%22%20fill-opacity%3D%220.08%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%22920%22%20cy%3D%22520%22%20r%3D%22170%22%20fill%3D%22%2322d3ee%22%20fill-opacity%3D%220.06%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M140%20430h920%22%20stroke%3D%22url(%23line)%22%20stroke-width%3D%2212%22%20stroke-linecap%3D%22round%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M220%20360c120-150%20260-220%20380-220s260%2070%20380%20220%22%20fill%3D%22none%22%20stroke%3D%22%2367e8f9%22%20stroke-opacity%3D%220.5%22%20stroke-width%3D%2216%22%20stroke-linecap%3D%22round%22%20stroke-linejoin%3D%22round%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M300%20398h600%22%20stroke%3D%22%2394a3b8%22%20stroke-opacity%3D%220.24%22%20stroke-width%3D%228%22%20stroke-linecap%3D%22round%22%2F%3E%0A%20%20%3Ctext%20x%3D%22220%22%20y%3D%22510%22%20fill%3D%22%23e2e8f0%22%20fill-opacity%3D%220.7%22%20font-family%3D%22Arial%2C%20Helvetica%2C%20sans-serif%22%20font-size%3D%2238%22%20letter-spacing%3D%228%22%3EINSIGHTS%3C%2Ftext%3E%0A%3C%2Fsvg%3E)
Free to Run, $410 Million to Build: Recursive's Flint Re-Prices Cloud Inference
Recursive's Flint, a 39.8-billion-parameter mixture-of-experts coding model released on October 21, 2026 under Apache 2.0 with no revenue gate, runs at more than 125 tokens per second on an M4 Pro MacBook Pro and was built in under four months with help from the company's automated research system. The leaderboard is the least important part: an unrestrictively licensed model running on hardware the developer already owns turns the price of an agentic coding token into a rounding error, while the same lab signed a $410 million compute deal with Amazon Web Services to build it. On-device inference relocates AI spending rather than removing it, so recurring per-token revenue migrates to the device while the lumpy training bill stays in the cloud and the fixed-cost layer in between gets squeezed first. We read that as value migrating toward compute, distribution and memory and away from the model layer, with CoreWeave carrying the clearest exposure and Microsoft the clearest transmission into a reported income statement.
Tesla's AI capex test failed: $4.7B of quarter spending didn’t match the market’s appetite—and cash generation weakened
In Tesla’s quarter ended Sep. 30, Tesla reported $4.7B of capital spending—well below the Street’s ~$7.9B expectation—and produced only $1.53B of free cash flow. At the same time, operating income fell to $718M as total operating expenses jumped $1.35B year over year and automotive regulatory-credit revenue dropped sharply, tightening the company’s ability to act as a high-conviction, open-ended “AI capex buyer.”
Uber Cut 10% of Its Support Staff in July and Cited AI. Then It Gave 30,000 Employees an Agent
Uber Technologies put ChatGPT Work in the hands of more than 30,000 employees on Oct. 22, 2026, seven weeks after cutting roughly 3,300 corporate roles and three months after cutting 10% of its customer support staff explicitly to \"embrace artificial intelligence.\" The headcount math is therefore already running, but not where the headlines assume: in the first half of 2026 cost of revenue rose just 3.6% while trips grew 18%, while general and administrative expense rose 30.7% and research and development 20.5%. Agentic software is already deflating the cost of serving a ride and doubling engineering output, and it is currently inflating the cost of running the company. The bill for 30,000 seats of agent access is metered in tokens that Uber Technologies has not published, which is why the Nov. 4 third-quarter print, not the October announcement, is the first real test.
Warhound’s $200M bet makes “robot dogs” a new defense unit-price race—and it pressures prime integration margins
Warhound emerged from stealth with a $200M seed round to build genetically enhanced military working dogs and outfit them with Anduril cybernetic/AI gear. The funding matters less for the round size than for what it implies: the Pentagon is now funding a cheap ground “manning substitute” while pushing more sensing/protection software down to component-level suppliers—changing how primes earn on integration.
2026-10-22
Anthropic's $80.7 Billion Loss and Amazon's $53.4 Billion Gain Are the Same Trade: the S-1's Real Number Is $518 Billion of Fixed Commitments
Anthropic's draft IPO prospectus, as reviewed by the Wall Street Journal and Reuters, shows a company that turned $1.1 billion of first-half 2025 revenue into $16.3 billion in the first half of 2026 while booking an $80.7 billion net loss — and, more usefully, a revenue book in which two unnamed customers produced about a fifth of sales while $518 billion of compute commitments, roughly 80% of it payable regardless of usage, run for seven to ten years. The cost side is locked and the revenue side renews, so the $2 trillion marketing price and the $1.5 trillion price investors now discuss are a renewal schedule rather than a growth debate. The same private marks drive Amazon's $53.4 billion quarterly gain and Alphabet's $98.0 billion of other income, and the debt stack financing the buildout — CoreWeave carries $51.6 billion of it — is where a single counterparty decision would surface first.
BlackRock’s new robotics ETF quietly validates the “integrators win” view—at least in today’s public-market pricing
BlackRock’s iShares Future Robotics and Related Technologies Active ETF launched on Oct 20, 2026 and already looks less like a bet on humanoid starring roles and more like exposure to the companies that build, automate, and industrialize physical systems. The initial portfolio weights skew toward industrial automation, machine-vision, and factory-enablement names, giving investors a public benchmark to stress-test private “embodied AI” valuations.
Claude Duet Reads 6,000 Apps but Cannot Move Your Money: Anthropic's Pre-Filing Agent Sends the Toll to Plaid and Visa
Anthropic's Claude Duet, launched Oct. 21, 2026 as a US-only research preview days before its public S-1 is expected to land, turns a chat window into an always-on agent that reads email, calendars and business ledgers, then pays merchants through Stripe Link or Shop Pay. The asset it assembles is not a better model but a permission graph: more than 6,000 connectors, a live Intuit integration and a Plaid bank rail that ChatGPT and Meta's Muse also rent. Because Anthropic blocks transfers between accounts and securities trading, banks capture nothing while the aggregator and card networks capture the toll, which is why Visa and Nvidia sit on the winning side of this launch and the software an agent merely reads, Intuit and HubSpot, have already been repriced.
AI Pentesting Went Commercial: 173 Tenants, One Stolen Secret, and a Fourth-Party Risk Nobody Repriced
Microsoft named Storm-0940, a China-aligned group, as the actor that used ARTEX — an open-source AI penetration-testing framework — to compromise ConnectWise's Cloud Backup service and attack hundreds of organizations worldwide. The mechanism was not a stolen password: attackers reused a vendor service-principal secret, and the agentic framework industrialised it into 2,250 tenant sign-in attempts at up to 2,900 a minute, with 173 Microsoft365 tenants confirmed compromised and privileged access inside three hours. The exposure now sits in the MSP layer that distributes trust and is priced by nobody — cyber insurance rates fell 4% in Q3 2026, a thirteenth straight quarterly decline, while identity vendors absorb the security budget and MSP software decelerates to mid-single digits. Detection and identity governance get paid for the new risk; the layers that absorb it keep getting priced for a benign world.
DeepSeek's $15 Billion Round Is the Toll Booth: One Domestic Chip Order Is Worth $5.6 Billion While NVIDIA Stays Shut Out
DeepSeek's roughly ¥100 billion ($14.98 billion) round — reported on Oct 21-22, 2026, one day after Washington published an order removing DeepSeek from federal civilian procurement and IT systems — was not a confidence event. The money came from inside the restricted perimeter: Tencent, CATL, GigaDevice, Midea, Meituan, Geely and a state AI fund, wired into the founder's partnership with no voting rights, so the check is procurement financing for compute that cannot be bought abroad. At least 160,000 Huawei Ascend 950DT accelerators for a gigawatt-scale Inner Mongolia campus come to roughly ¥40 billion ($5.6 billion) at list prices that rose 20-50% in two months, while NVIDIA's realized H200 sales to China stayed under 1% of data center revenue. For US equity investors this is a supply-chain signal, not a China-tech headline: the redirected capex is landing on Chinese silicon, memory and power, and NVIDIA's own filing says the foreclosure helped competitors build ecosystems that challenge it worldwide.
The 30-Year Fell to 5.32% While Oil Hit $112: The Bond Market Quit Pricing Inflation and Started Pricing AI Credit
On Oct. 22, 2026 the 30-year Treasury yield fell to 5.317% — its lowest since Sept. 23 — in the same session Brent crude jumped 2.6% to $111.64, a pairing that used to be read as an inflation scare. It wasn't: five- and ten-year breakevens never left 2.4%, so the whole long-end rally was real yield, set by a market now pricing AI capex as a credit risk rather than an inflation risk. Meta Platforms's five-year credit default swap touched a record 119.3 basis points the same afternoon, the $13 billion 20-year reopening cleared at 5.469% — the highest yield for that maturity since 2020 — and Oracle now runs negative $28.7 billion of trailing free cash flow. The trade that follows is long the AI supply chain that converts capex into cash, and short the AI structure that has to refinance itself at 15%.
Meta's New Terms Make Publisher Content Free by Contract — and Publishers and Advertisers Just United to Object
From October 30, Meta's rewritten commercial terms let it crawl any website tied to a business account, with no opt-out and no contractual promise to honour robots.txt. On October 22 the Movement for an Open Web — a group whose members include digital advertisers and publishers — asked the European Commission and the UK's Competition and Markets Authority to suspend the change, and publisher groups filed in Brussels, Berlin and London in parallel. The alignment is the real signal: one clause sweeps up both the content a publisher sells and the product data an advertiser keeps, and Meta is offering both constituencies the same trade — a licence for free in exchange for a relationship they cannot walk away from. Because the CMA cannot use its Digital Markets Act powers against Meta before Friday, investors should price the licensing reset rather than the fine.
One Stolen OAuth Secret Put 1 Million SMBs in Scope: The Chokepoint Moved to the IT-Services Channel
Attackers turned one compromised application credential at ConnectWise's Cloud Backup service into a cross-tenant key for Microsoft 365: Huntress confirmed 173 unauthorized sign-ins, Microsoft counted more than 2,250 tenants attempted, and one ConnectWise app self-granted Global Administrator inside a partner tenant. Because that credential sat behind ConnectWise's 44,000 partners serving more than 1 million SMBs, the blast radius was set by vendor reach rather than by a software bug, and Microsoft's own threat intelligence attributes the operation to a China-aligned group running an agentic AI framework against hundreds of organizations. The remediation budget now runs through three lines: identity governance at Microsoft, backup and recovery at Commvault and Rubrik, and channel-delivered detection at SentinelOne and CrowdStrike. The closest listed peer to the breached vendor, N-able, is de-rating instead — the chokepoint itself is privately held, so investors own the fixes, not the victim.
OpenAI's Models Found a 9.2-Critical Flaw in 70% of the Web. Nobody Paid for It, and the Money Went to the Edge
A frontier lab found two critical PHP remote-code-execution flaws, ran the two-month global coordination that fixed them, and shipped a free version checker to every developer with a Codex seat. What changed is the plumbing: AI has made discovery free, so the scarce input is now triage and enforcement, which is where the money goes \u2014 to web application firewalls, exposure platforms and the sponsors who fund PHP's maintainers, not to the bounty desk. For an IPO-bound lab, a coordinated disclosure program is a funnel and an unpriced liability; for investors, the split between platform security and pure-play scanners is now visible in the filings.
2026-10-21
Anthropic's $80.7 Billion First-Half Loss Is Mostly a Mark on Amazon's Own Paper. The $518 Billion It Owes Is the Number That Bites
Anthropic's first-half 2026 net loss of $80.7 billion is roughly 89% a non-cash charge from revaluing the convertible notes that funded its compute buildout, leaving about $8.7 billion of real loss and a $55.7 billion mark-up on Amazon's Anthropic position in a single quarter. The economic story is not the write-down; it is the payment schedule underneath it. The prospectus discloses at least $518 billion of compute obligations over about a decade, roughly 80% of it non-cancelable or payable regardless of usage, against $20.28 billion of cash at the end of 2025. A $2 trillion November listing is therefore less a valuation verdict than the funding mechanism for a fixed schedule, which makes the credit cycle that finances AI capacity, not the model, the thing investors are actually underwriting.
Fireworks AI's $40 Billion Ask Values the Inference Software Layer at 40x Revenue — and NVIDIA, Its Supplier and Shareholder, at 17x
Fireworks AI is in early talks to raise at a $30–40 billion valuation, roughly double the $17.5 billion it marked in July on a run rate it disclosed as more than $1 billion. That puts inference-serving software at 30–40x revenue, against Modal's 15.8x set on October 14, fal.ai's above 18.8x, NVIDIA at 17.3x and CoreWeave at 10.5x enterprise value to sales. The catch is the margin structure: Fireworks' gross margin is estimated near 50%, below CoreWeave's 65.9% and NVIDIA's 75.0%, so the layer winning the software multiple looks more like a hardware reseller. The mechanism matters more than the multiple — private capital raised by the software layer is what funds the GPUs it buys, so every mark raised upstream converts into chip orders downstream, and the balance-sheet cost lands on the neoclouds.
Getty’s Chapter 11 turns the “AI broke the scarcity premium” story into a creditor fight over the image catalog
Getty Images’ prepackaged Chapter 11 restructures control of its rights catalog before any court-versus-contract debate about who can monetize (or withhold) image licensing after AI training. The filing also hardens a key market lesson: when a licensor can’t stay solvent, the value of licensed content becomes less about AI scarcity and more about deal terms, creditor approvals, and post-restructuring ownership.
Meta's $900 Billion Class C Is Not a Second AI Security — the Spread, Not the Listing, Prices the Buildout
Meta's board approved a new non-voting Class C share on October 21, 2026, to be distributed one-for-one to every Class A and Class B holder, with Class A trading as a two-for-one split once it lands in January 2027 — so the class raises nothing on day one and changes no shareholder's vote. What it changes is the menu: free cash flow turned negative at -$2.12 billion in the September quarter, buybacks were zero for a second straight quarter, and the 2026 capital spending guide was narrowed up to $135-145 billion, leaving roughly $56 billion more to fund between 2026 and 2027 consensus. The Founder Alignment Agreement staples Mark Zuckerberg's Class C to his Class B and frees only 2.5% a quarter for eight years, which makes the Class C discount to Class A a clean daily quote on the equity cost of the AI buildout — the leg credit markets already priced, with Meta's $27.3 billion Hyperion bond near 91 cents.
Nvidia Is Buying Europe's Certification Layer, Not Its Permits: 25,000 Lucid Vehicles, 20 Approved Robotaxis, and Mobileye Squeezed
On October 21 Nvidia put Bolt's European robotaxi programme on its own Alpamayo and Hyperion stack, extended the same architecture to Einride's trucks, and pulled Europe's tier-1 suppliers, OEMs and safety validators onto one reference design — while the only robotaxi service permit granted on the continent still covers 20 supervised vehicles in Madrid. The scarce asset in European autonomy is therefore not a fleet but an authorisation written against a stack, and Nvidia's move is to own the layer those authorisations cite. That squeezes Mobileye's licence model, leaves Lucid's 25,000-vehicle target unfunded, and forces Uber to fund three rival stacks at once — widening the gap between the platform that owns the evidence trail and the vehicle makers, operators and demand aggregators that own the customer.
OpenAI on Google Cloud Is a Capacity Play, Not a Partnership: Google Takes 1.5–3% of a $50B Run Rate as Azure's Exclusivity Premium Disappears
On October 21, 2026, OpenAI put its frontier models and Codex on Google Cloud's Model Garden, sold through Google Cloud Marketplace, with enterprises able to draw existing Google Cloud spend commitments down against those purchases. The money is not in the model: Alphabet's cloud segment grew revenue 82% to $24.8 billion in the June quarter while free cash flow turned negative on $44.9 billion of quarterly capital spending, and 2026 capex guidance now sits at $195–205 billion. Google is renting out a TPU fleet it has already financed, and the published Marketplace schedule keeps 1.5% to 3% of every transaction — which turns Anthropic's disclosure that 47% of its 2025 revenue flowed through cloud marketplaces into the most important undisclosed number in AI distribution. Microsoft keeps first-launch rights and OpenAI's $250 billion Azure commitment through 2032, but the exclusivity premium that once justified both expired in April 2026, and OpenAI's own infrastructure plan has been reset to roughly $600 billion through 2030 rather than $1 trillion.
$25 Million and One Signature: How OpenAI's President Ended Up on Both Sides of America's AI Rulebook
OpenAI president Greg Brockman signed the White House's voluntary AI safety accord on Sept. 29, 2026 and, weeks later, was reported to have given $25 million to Leading the Future, the super-PAC network that fights rules said to delay AI development - while his own company's policy chief publicly asks Congress for mandatory national safety standards. The two positions are not accidental: Brockman's money targets the instrument, while OpenAI's lobbying targets a single national framework it can shape and its competitors cannot. That combination lands at the worst possible moment for an IPO, because the confidential S-1 was submitted before either the accord or the donation became public, and because the Federal Trade Commission is now demanding the incident and marketing records that make safety claims testable. For public investors, the trade is not 'regulation good or bad for AI' - it is that OpenAI's own execution shortfall, a September run rate near $50 billion against roughly $70 billion signaled, does more damage to the AI capital complex than any AI bill does.
What to expect
Evidence-first notes with a visible point of view.
This section collects sharp takes on earnings, shareholder meetings, and market structure. Each new piece should make the thesis, the facts, and the implications obvious within the first few screens.
Expect direct analysis, not generic commentary.
Expect the data to be explicit and the argument to be easy to follow.