Amazon, Google, and Microsoft Funnel Money into Anthropic’s Coffers

Anthropic Brings SpaceX a Third of Its Revenue and Relies Heavily on Three Hyperscalers

Anthropic has become SpaceX’s largest customer, accounting for 34.5 percent of total revenue in the past quarter, according to SpaceX’s latest quarterly figures. A growing share of revenue for Elon Musk’s company now stems from its data center and computing infrastructure business for AI. This segment is growing rapidly, but current disclosures reveal it depends heavily on a single customer.

Anthropic filed its S-1 prospectus on Monday, aiming for a Nasdaq listing under the ticker ANT. The offering price and number of shares have not yet been determined. According to the prospectus, revenue in the June quarter rose 1,368 percent year-over-year. In the first half of the year, the company generated $16.3 billion in revenue. The net loss for the same period stood at $80.7 billion, largely driven by non-cash valuation adjustments on convertible bonds. The operating loss narrowed to $693 million, despite significant compute commitments.

Heavy concentration is evident on Anthropic’s side as well. According to the prospectus, almost half of its revenue comes from Amazon, Google, and Microsoft. Concerns are mounting among investors that the AI industry and the broader economy depend too heavily on just a few companies. If one of them falters, markets could be destabilized. In a client note, hedge fund Bridgewater wrote that concentration and acceleration are inherently risky: navigating them requires “more plain luck” to keep economies running.

Investor Kevin O’Leary called the Anthropic IPO a potential “hot trade” on X. He had declined his allocation in the SpaceX IPO because he would have had to hold the shares for 30 days. For Anthropic, he expects day-one tradeability. Anthropic, however, has not confirmed whether all allocations can be sold immediately, as underwriting banks can impose restrictions. O’Leary generally recommends betting on SpaceX, Anthropic, and OpenAI simultaneously, rather than trying to pick a single AI winner. → Business Today, Benzinga, Semafor

Synthszr Take: O’Leary’s advice to bet on SpaceX, Anthropic, and OpenAI all at once does little to diversify risk. Over a third of SpaceX’s revenue comes from Anthropic. Nearly half of Anthropic’s revenue comes from Amazon, Google, and Microsoft. Musk’s AI infrastructure is thus tied to the cloud budgets of three corporations he doesn’t control. A basket of these names holds a single bet, simply purchased multiple times; Bridgewater’s “plain luck” is already priced in.

CrowdStrike Reports Chinese DeepSeek Agents Attacked 200 Targets, Breaching Nine

According to a report by CrowdStrike, China-nexus threat group VAPOR PANDA attacked more than 200 organizations across at least 14 countries in 11 days using the AI-driven attack framework ARTEX. For nine of them, the security firm confirmed post-intrusion activity, including the deployment of the BLOODALCHEMY backdoor. CrowdStrike noted that ARTEX ran primarily on the Chinese language model DeepSeek, complemented by a second, commercial subscription model that CrowdStrike could not identify. The framework coordinates a primary agent with subordinate planning and execution agents, pursuing multiple attack paths in parallel and pivoting to alternatives when one fails. → Unite.AI

Synthszr Take: 1.4 billion input tokens for a single intrusion is primarily an invoice, and at DeepSeek prices, it’s a small one. The attacker scales with their token budget, while most defenders are still calculating in analyst hours. An operator who only intervenes at key decision points can target more than 200 organizations in 11 days. A security team reviewing every alert by hand is structurally set up to lose this race. Defense must therefore run agentically as well (which CrowdStrike, naturally, is happy to sell).

Ex-Palantir Executives Aim to Replace AutoCAD and Revit with AI

California-based startup Dream Technologies has launched with a $17 million seed round, according to SiliconANGLE. The round was led by Andreessen Horowitz, with participation from SV Angel, XYZ, and Zigg Capital, among others. Dream was founded last year by former Palantir executives Andy Elder and Arjun Mathur. The company developed AI models that progressively generate floor plans and detailed component dimensions from site data and project requirements. According to the company, hundreds of design iterations can be produced in minutes. The models were trained on data spanning approximately 100 million square feet of US real estate, ranging from commercial properties to single-family homes. → SiliconANGLE

Synthszr Take: Dream assumes liability for its plans, and that decision defines its business model. As the Architect of Record, the startup competes directly with architectural firms for design fees, bypassing the likes of Autodesk entirely. Liability is a moat no software vendor wants to touch. However, it scales poorly: each new state requires licensed architects, insurance, and compliance with local building codes. Ten projects representing $20 million in volume make for a solid pilot. Andreessen Horowitz is betting that compute can replace planner hours more cheaply than liability costs.

Google’s Suncatcher Satellite Lasts Just 20 Minutes in Orbit

Google’s prototype satellite from Project Suncatcher is operational and communicating with ground control, TechCrunch reports. It launched earlier this month alongside more than 100 other spacecraft aboard a SpaceX Falcon 9. Onboard are four Tensor Processing Units (TPUs), powered by a solar array that generates up to 1 kW. Due to this tight energy budget, only a single TPU can operate at a time for about 20 minutes, complicating communications between orbit and ground. During the initial test, the pre-installed Gemma 4 model from Google DeepMind composed a haiku about space before transmitting telemetry on orbital conditions around the satellite. → TechCrunch

Synthszr Take: Four TPUs on board, power for only one, and even that for just twenty minutes. This reduces the business model of orbital data centers to a single metric: kilowatts per kilogram of launch mass. Every additional TPU demands solar array area and cooling, both of which can only be launched economically once a fully reusable Starship is operational. Google’s expansion beyond 2027 is therefore dependent on Musk’s rocket (Alphabet conveniently holds a stake in SpaceX). Until then, orbital computing remains a shift-based operation with haiku demos.

Spotify Plans Surcharge for AI Remixes Starting in 2027

Spotify is planning an AI tool for the first quarter of 2027 that will allow paying subscribers to create custom remixes and cover versions of songs, according to the Financial Times. The company is considering an extra charge of around $4 to $5 per month for this add-on feature. Sources tell the publication that a public debut could align with the Grammy Awards in February, though the schedule and pricing remain unconfirmed. Users would be able to transform a folk track into a dance mix or have one artist appear to cover another’s song. Royalties from fan remixes would flow to the original artists, and remix streams would be credited to the original song on Spotify’s charts. → iPhone in Canada

Synthszr Take: Spotify’s margins have always suffered because record labels capture the lion’s share of every subscription dollar. The remix tier layers a second variable cost on top of this licensing burden: compute for every generated song. Under a flat four-to-five-dollar monthly fee, the most active fan becomes the most expensive. Spotify will likely have to implement usage quotas—credits instead of an unlimited flat rate. Only then will we see whether fans are willing to pay to tinker with songs, or merely to listen.

Bank of America Spots Stock Market Bubble Signal Reminiscent of Dot-Com Crash

Bank of America analysts have identified an equity market pattern last seen right before the dot-com bubble burst, Business Insider reports. The bank described the signal as “bubble-like,” referring to an explosive rally in a single segment of the market that analysts consider historic. According to the bank, a comparable pattern emerged in the lead-up to the crash in 2000. → Business Insider

Synthszr Take: In 1999, capital poured into Super Bowl commercials and domain names; today, it flows into data centers, silicon, and power purchase agreements. That offers less comfort to shareholders than it might seem. Global Crossing also laid real fiber back then and still went bankrupt. The cables remained in the ground, making streaming cheap years later. If this rally bursts, investors pick up the tab, while companies building atop the infrastructure will snap up computing capacity at firesale prices.

CrowdStrike CTO Warns: Shadow AI Is Outgrowing Enterprise Control

CrowdStrike’s Field CTO for the Americas, Cristian Rodriguez, warned on the podcast The Deep View Conversations that while AI agents accelerate employee workflows, they also introduce unprecedented risks to sensitive corporate data. The conversation was recorded at CrowdStrike’s Fal.Con 2026 conference. The central discussion focused on how enterprises handle Shadow AI—AI tools introduced by employees without IT approval. Rodriguez argued that defenders must now respond at machine speed. As an example of purpose-driven agents, he pointed to an incident involving OpenAI and Hugging Face. → The Deep View

Synthszr Take: IT approval is the slowest link in enterprise AI adoption, and employees have long since learned to bypass it. Every self-installed agent accumulates access rights that were never officially granted. The data exposure gap expands in tandem with productivity; banning these tools would only cripple output. (Even Rodriguez runs his schedule with a Slackbot.) For CrowdStrike, this unlocks a whole new market: agent governance delivered as continuous surveillance, replacing the traditional approval workflow rather than simply accelerating it.

Addy Osmani: Autonomous Verification Is What Truly Scales AI Agents

Google developer Addy Osmani argued in his Substack article “Agentic Verification” that AI agents produce significantly better outcomes when they can autonomously review their own work. The essay originally ran about two weeks ago in his “Elevate” newsletter and was recently featured again in Substack’s personalized weekly digest. Osmani’s core thesis is that agents write code “far better on average and at scale.” He specifically attributes the “at scale” qualifier to Agentic Verification—verification steps executed directly by the agent itself. In his view, agents make verification cheap, and only then can quality be sustained across high volumes of tasks. → Substack

Synthszr Take: Agents make verifying code cheap, but verifying claims about agents remains expensive. Osmani’s post reappears in the Substack digest 15 days later, sans benchmarks, sandwiched between Kevin Kelly and a piece on recursive self-improvement. The algorithm optimizes for engagement and likes; nothing is rigorously validated. As a result, the bottleneck shifts from writing code to filtering assertions. Teams procuring agent workflows need a curated verification layer, or they risk adopting whatever thesis has the catchiest pull quote.

Stanford Team Sets Out to Formally Verify Software

Stanford computer scientist Chuyue Sun and her team have developed an AI tool that completed a verification project in under half a day for less than $500. Three highly skilled specialists had previously spent five months performing the same work manually. The benchmark test case was a section of code responsible for encrypting chats in the Signal messaging app. Sun published the results in a recent research paper. They represent some of the earliest empirical evidence bringing renewed attention to the historically niche field of formal verification.

The methodology mathematically proves that software operates without bugs. To do this, developers must translate the intended behavior of their software into precise specifications and express the code as verifiable formulas. The overhead is so steep that companies typically stick to traditional testing suites. Formal verification has historically been reserved for mission-critical systems, such as those at NASA or Apple.

Sun joined &Truth in August, a startup that went public with its launch on October 29, 2026. Established as a Public Benefit Corporation, the company raised $70 million in seed funding. It plans to build a universal verification engine capable of parsing code in any programming language, validating its security and correctness through independently checkable proofs. According to the company, AI accelerates the proof generation process by a factor of 100 or more, with proprietary automation adding another 100x speedup. A “Specification Factory” is designed to translate business requirements into formal specifications at scale. The technical stack builds on the open-source Lean proof assistant, complemented by the Mathlib mathematics library with over two million lines of code, as well as CSLib. &Truth also intends to release its core libraries as open source.

&Truth was co-founded by Clark Barrett, Stanford professor and lead of the reasoning engine cvc5; Ilya Sergey, who established the VERSE Lab at the National University of Singapore; and Chris Hsu, former Managing Director at Citadel. The company is based in Palo Alto and Singapore. Sergey draws an analogy to structural engineering: bridges and skyscrapers don’t collapse because structural calculations are completed prior to construction, and software engineering should adhere to the same standard.

Competition is taking shape concurrently. Harmonic, co-founded by Robinhood CEO Vlad Tenev and backed by $295 million in funding, has expanded its mathematics agent Aristotle to support C code and hardware designs in Verilog. The platform is currently being piloted with prospective clients including chipmakers Tensordyne and Nvidia, whose venture arm is an investor. A week prior, Anthropic and formal verification startup Theorem announced they had leveraged AI to identify defects in open-source software that had evaded trillions of test cycles. Alongside this, they rolled out an extension for Claude aimed at assisting developers with verification. → Wired, unite, Puck

Synthszr Take: Three experts, five months, under $500: this calculation dismantles the hourly billing model of software audits. Audit firms are suddenly selling a service that costs less to deliver than their kickoff meeting. Formal verification is transitioning from an exclusive NASA privilege into a baseline requirement for financial software. The only expensive piece left is what the machine cannot automate: articulating precisely what the code is supposed to do. That is why &Truth is building a Specification Factory—and that is where the new pricing model will be written.