AI Supply Chain Breach | AI Field Notes #84

AnalysisTwo dollars in, six dollars out, per million tokens: that is what Elon Musk's AI company is charging for Grok 4.6, released August 12. It scored 61 on the Artificial Analysis Index, an aggregate benchmark of model quality, drawing level with OpenAI's best current model while costing less to run. Token prices are the meter developers actually watch, since they decide which model gets wired into a product and which stays a demo. Frontier-level quality at a discount pressures every rival's pricing page, and it turns the choice of default model into a monthly negotiation rather than a settled bet.

AnalysisRotate your keys. That is the advice landing on security teams at roughly 2,500 companies, among them AWS, Cisco, Salesforce and Samsung, after researchers at CloudSEK disclosed on August 11 that two tainted versions of LiteLLM, a widely used open-source gateway that routes apps to different AI models, quietly harvested secrets from about 434,000 build pipelines. The attackers, a group tracked as TeamPCP, poisoned a scanning tool inside LiteLLM's own release process, so the malicious code shipped through the front door. Cloud keys, repository tokens, SSH keys and Kubernetes secrets were all in reach. Removing the package does not undo stolen credentials.

AnalysisFounded barely 18 months ago, Lovable just raised 400 million dollars at a 13.3 billion dollar valuation, close to ten times its Series A, on the promise that people can build working software by describing it in plain English. The company counts more than 5,000 customers and sits at the center of the vibe coding wave, where the person shipping an app never opens a code editor. That valuation is a bet that the next generation of software gets made by people who cannot read the code underneath. Whether those apps survive contact with real users is the open question.

AnalysisTen employees, ten million dollars in annual revenue, and now a valuation that jumped almost tenfold in under a year: Blacksmith raised 45 million dollars on August 12 at a 550 million dollar valuation, betting that the bottleneck in software has moved from writing code to checking it. When coding agents generate pull requests faster than any human can review them, the scarce work becomes automated testing and validation, which is what Blacksmith sells to more than 5,000 customers. The round is a wager that the market for catching machine-written bugs grows as fast as the machines writing them.

AnalysisChip companies do not usually give away the software that sells the chips, which is what makes Nvidia's plan notable. The Information reported on August 11 that Nvidia is training Nemotron 4, an open-weight model family whose largest version targets at least a trillion parameters, free for any company or country to run. The logic runs in Nvidia's favor: more capable open models mean more demand for the hardware that trains and serves them, and Nvidia has budgeted about 28 billion dollars in cloud spending through 2031 to feed the effort. Chinese labs already ship models past that parameter scale, so this is catch-up dressed as generosity.

AnalysisThe developers OpenAI most wants building on its tools had been stuck in a browser tab, and on August 11 that changed: the ChatGPT and Codex desktop apps arrived on Linux, covering Ubuntu 24.04 and 26.04, Debian 13, and Fedora 43 and 44 on both Intel and ARM chips. Linux is where most servers and a large share of working engineers live, so shipping a native app signals where OpenAI's paying technical audience actually is. Codex, the company's coding agent, now sits one keystroke from the terminal it operates in, closing a small gap that annoyed the exact users OpenAI courts.

AnalysisA billion people now open Google's Gemini app every month, a milestone Sundar Pichai announced on August 11 and called the company's fastest-growing product. The texture of that use matters more than the round number: 63 percent of interactions come through voice, and users generate more than 150 million images a day, which points to a public treating AI as something you talk to and draw with rather than type at. It arrives weeks after OpenAI's ChatGPT reported the same billion-user mark, so consumer AI now has two products at genuine mass scale, funded by very different businesses.

AnalysisFor the first time, the boxes running AI models earn Foxconn more than the boxes people hold in their hands. The world's largest contract manufacturer posted record quarterly net profit of 59.98 billion Taiwan dollars, about 1.86 billion US dollars, up 35 percent from a year earlier, with earnings per share climbing to NT$4.27, Taipei Times reported on August 13. Its cloud and networking division, which builds AI servers, crossed 51 percent of revenue, overtaking the iPhone assembly that defined the company for two decades. Foxconn is now, in the numbers, an AI infrastructure company that still happens to make phones.

AnalysisThe attackers barely touched a keyboard. Near-autonomous AI agents ran reconnaissance and launched simultaneous break-in attempts against Taiwan's nuclear safety regulator, in an intrusion The Register reported on August 12 and tied to China. What makes it a marker is the automation: the software scoped the target and coordinated the assault at machine speed, with a human somewhere approving rather than typing. Defenders now face attackers who can probe thousands of doors at once and never tire. The same agent tooling being sold to speed up engineering is being pointed, quietly, at critical infrastructure.

AnalysisRenting compute from a rival is an odd move until you are out of your own. Together AI, which sells developers access to run AI models, agreed to a 240 million dollar multiyear deal on IBM Cloud, deploying roughly 2,000 of Nvidia's Blackwell HGX B300 systems starting in the first quarter of 2027, according to The Register on August 11. The tell is that even a company whose whole business is inference capacity, the cost of running a model for users, is now leasing someone else's data center to keep up. Compute scarcity is no longer a talking point; it is a purchase order.

AnalysisThe year is not two-thirds done, and AI-attributed layoffs in the United States have already reached 205,000 workers, matching the full-year 2025 total in under eight months. In more than half of the tracked cuts, 170,945 jobs across 173 companies, employers named AI, automation, or machine learning as a driver, with the damage concentrated in customer service, data operations, entry-level software, and finance back offices. The pattern underneath is quieter than the headline: companies are thinning the rungs where people used to learn the job, cutting the trainee roles that once fed the senior ones. The ladder is losing its bottom steps.