August 31, 2026
Nvidia just put $3.5 billion into MediaTek, the Taiwanese chipmaker best known for phone processors. On its face, that's an odd move for the company that already dominates AI training hardware. But TechCrunch's reporting gets at the real story: every major Nvidia customer — Amazon, Google, Microsoft, Meta — is quietly building its own AI chips to cut Nvidia out of the bill of materials. Nvidia's answer isn't to out-engineer them. It's to make sure that whoever wins the custom-silicon race still needs Nvidia's networking, software stack, or now, a manufacturing partner it has a financial stake in. For business buyers, this matters less as a stock story and more as a signal about where AI compute costs are headed. If Big Tech's custom chips actually ship at scale, the price of running large models could fall faster than most budgets currently assume. That's good news for anyone paying per-token or per-seat for AI features today. But it also means the ground under 'which AI vendor should we bet on' keeps shifting, which is exactly why teams are better off choosing platforms that abstract away the underlying model and hardware rather than locking into one vendor's roadmap — something worth weighing when you're comparing a no-code app builder against building directly on a single AI provider's stack.
Sony Music and Warner have sued Anthropic, and according to TechCrunch, this complaint goes further than earlier AI copyright suits by explicitly alleging illegal piracy, not just unlicensed training. That's a meaningfully different legal argument, and if it holds up, it raises the potential liability for any company whose foundation models were trained on scraped material with unclear provenance. Businesses building products on top of large language models should treat this as a preview, not a one-off. Every AI vendor that trained on the open web has some version of this exposure, and the outcome of this case could shape licensing costs, model availability, or even which vendors survive with their current business models intact. I don't think this ends with Anthropic; I think it's the first of several tests of whether 'we didn't know the data was pirated' holds up as a defense. Companies picking AI tools for content generation, marketing copy, or media search should start asking vendors directly about training data provenance rather than assuming it's someone else's problem.
Two smaller stories underline the same point. Circleback, the meeting note-taker, just added a free tier and new plans starting at $14 a month, per TechCrunch — a classic land-grab move once a market gets crowded with competitors offering similar transcription and summary features. Meanwhile, Clipto, a three-year-old startup that searches terabytes of video using AI, just hit a $250 million valuation on the back of $15 million in ARR and, notably, actual profitability before its latest raise. What connects these two is that neither is selling a foundation model — they're selling a workflow wrapped around one. Clipto's profitability at a relatively modest ARR is the more interesting data point here: it suggests that narrow, well-executed AI tools solving one specific search-and-retrieval problem can be sustainable businesses without needing hyperscaler-level funding. That's a useful reminder for any team evaluating whether they need a sprawling, expensive AI platform or something narrower and cheaper that just does one job well, whether that's meeting notes or analytics and reporting built into tools they already use. Circleback's free tier, meanwhile, is a bet that distribution now matters more than differentiation — get enough people using the free version and monetize the ones who need more, a strategy that only works if the underlying product margins allow it.
Which of these developments actually changes your buying decisions this quarter: the chip economics, the copyright risk, or the pricing pressure on AI point solutions?
Sources