DeepSeek builds chips, OpenAI buys loyalty, and AI crime gets fact-checked

DeepSeek designs its own chip, OpenAI and Anthropic spend $800M/yr on startup credits, and the 'first AI ransomware attack' was more human than headlines claimed.

The big picture

Two themes dominate this batch: the AI business model is getting weird (compute giveaways, IPO pressure, Chinese models eating market share), and the “AI did X autonomously” headlines keep needing significant asterisks once you read past the first paragraph.

The compute credit arms race and what it signals

OpenAI and Anthropic are giving away compute credits at a scale that sounds absurd until you think about it: up to $800 million per year combined at Y Combinator alone, with individual offers exceeding $3 million per startup, according to The Decoder. The cloud providers are piling on too.

This is ecosystem lock-in 101, but the timing is telling. Both companies are eyeing IPOs and need to show developer adoption at scale. Free credits convert into reference customers, then (they hope) into paid contracts. The uncomfortable math: if you’re giving away $800M to win customers who might spend $200M when the credits expire, you’re betting heavily that switching costs will be high enough to keep them. For you as a developer, this is free money — take it. Just don’t architect your entire system around a single provider’s proprietary features while you’re spending it.

Meanwhile, Chinese models already account for over 30% of traffic on OpenRouter and are gaining ground specifically because of the price gap with OpenAI and Anthropic, per The Decoder. And DeepSeek just announced it’s designing its own AI chip, which would reduce its dependence on Nvidia hardware it can’t reliably import due to export controls (The Decoder). If DeepSeek gets viable custom silicon, the cost gap widens further. The free credits from OpenAI and Anthropic start to look less like generosity and more like urgency.

The autonomous AI threat: messier than the headlines

Last week’s story about the “first AI-run ransomware attack” was real, but the framing was doing a lot of work. New reporting from TechCrunch clarifies that a human still selected the victim, built the infrastructure, and handed over stolen credentials — the AI handled the technical execution steps. That’s meaningful (AI as a force multiplier for less-skilled attackers), but it’s not the fully autonomous criminal AI the initial headlines implied.

The distinction matters for how you think about threat modeling. An AI that lowers the skill floor for executing an attack is genuinely concerning. An AI that selects victims, sources credentials, and orchestrates everything end-to-end with no human in the loop is a different category of threat entirely. We’re not there yet, and security coverage that blurs that line doesn’t help anyone build realistic defenses.

On the defensive side, a startup called Savi just raised $7 million in seed funding and is launching an iPhone and Android app to detect AI-powered scams, specifically the kind where synthetic voices impersonate family members demanding ransom (TechCrunch). Consumer-facing scam detection is an underfunded problem given how fast voice cloning has matured, so this is a space worth watching.

Autonomous systems in the physical world

Forterra has deployed over 100 autonomous ground vehicles in Ukraine, making them the first American autonomous vehicles in active combat (TechCrunch). This is a significant real-world stress test for autonomous vehicle software in adversarial, unstructured environments — conditions that make a San Francisco robotaxi run look trivial by comparison.

Ars Technica’s feature on general-purpose robot autonomy surveys researchers and founders on where the field is heading for workplace and home applications. The honest summary: the hardware and manipulation capabilities are advancing fast, but reliable autonomy in unstructured human environments remains the hard part. Worth reading if you’re building anything that touches robotics or physical AI — the researchers are candid about timelines in ways that press releases aren’t.

The AI profit timeline reality check

Apollo’s chief economist Torsten Slok is pushing back on Wall Street’s optimistic AI productivity timeline. His argument, covered by The Decoder: regulated industries like healthcare, banking, and pharma face compliance overhead and process overhaul requirements that could push meaningful margin gains years out, not months. If the five-year scenario is closer to reality than the five-month scenario, a lot of AI-adjacent stocks are priced for a world that won’t arrive on schedule.

For developers building in those regulated verticals, this is less a warning and more a confirmation of what you already know from dealing with procurement, legal, and compliance cycles. The technology moving fast and the organization adopting it moving slowly are two separate clocks.

Quick hits

SK Hynix is pursuing a multibillion-dollar U.S. IPO, riding AI-driven HBM memory demand — another datapoint on how the AI infrastructure boom is minting hardware winners. TechCrunch
Solos launched the AirGo A6 smart glasses at 19 grams (no camera, voice AI only), about a third the weight of the new Meta Glasses — a different design philosophy that prioritizes wearability over compute. The Verge
Two MIT Technology Review pieces on AI architecture foundations for enterprise and autonomous enterprise systems are sponsored content territory, but the framing around agentic risk and investment durability reflects real conversations happening in large IT orgs right now. MIT Technology Review