Future Technology[Daily Briefing] DeepSeek builds its own chip 🧠, ransomware that thinks 🔒
DeepSeek is building its own chip to stop leaning on Nvidia
Three separate sources confirmed on 7 July that DeepSeek, the Hangzhou startup that rattled the AI world last year, is designing its own silicon. The interesting part is what it is for. This chip is built for inference, the stage where a trained model actually answers you, not for the giant one-off job of training the model in the first place.
That focus is the whole story. Training grabs the headlines, but inference is where most of the day to day compute cost quietly lives, because a model answers millions of questions over its life and each one costs a little power. A cheaper in-house inference chip could change DeepSeek's unit economics in a way that shows up every single day.
There is a bigger board behind this too. US export controls block China from buying Nvidia's best chips, so DeepSeek runs on a mix of older Nvidia parts and Huawei's Ascend processors, and Beijing keeps pushing its champions to build domestic alternatives. DeepSeek has a reported 7.5 billion dollar war chest, so it can fund a multi-year hardware effort. The project is roughly a year in, with no tape-out date yet, and Nvidia shares dipped about 1.6 percent on the news. Not everyone is worried: analyst Richard Windsor reckons the chip may struggle to win customers outside China, which would keep its early impact mostly at home.
Why it matters: the next fight in AI hardware is not about who trains the biggest model, it is about who can serve one cheaply at scale. If DeepSeek pulls off a low-cost inference chip, it chips away at the single most profitable corner of Nvidia's business and gives China a working answer to the export ban. That is a slow burn, not a knockout, but it points at where the money moves next.
Read the full breakdown on Future Technology
Quick hits
- Etched lands a 5 billion dollar valuation on inference bets. The chip startup says it has already signed 1 billion dollars in inference-chip contracts, another sign the money is chasing the answering stage, not just training. Why inference is the new battleground (3 min read, source: TechCrunch)
- The first ransomware that improvises just showed up. Researchers documented JadePuffer, described as the first agentic ransomware, running an AI agent that retries failed steps and adapts instead of following a fixed script. Static defences assume the attacker sticks to a known playbook, and this one does not. How AI is reshaping ransomware (4 min read, source: crescendo.ai)
- Reflection AI switches on a 6.3 billion dollar compute lease. It activated GB300 chips at SpaceX's Colossus 2 site in Memphis through 2029 to train American open-weight frontier models. These lease deals are becoming the real scoreboard for who can train at the frontier. Inside the Colossus 2 deal (3 min read, source: aiapps.com)
- Amazon is quietly designing its own AI chips. Reports say custom silicon is coming for Echo and Fire TV, alongside a reported 25 billion dollar bond raise for AI infrastructure. Big tech keeps deciding it would rather own the chip than rent it. The chip land grab, explained (3 min read, source: techstartups.com)
- We just passed 390 gravitational wave detections. Astronomers released the biggest catalog ever on 1 July, adding 161 new black hole collisions. Ten years ago the count was zero. That pace is the real wonder here. See the record catalog (4 min read, source: sciencedaily.com)
Tool of the day: a hardware security key
In a week when ransomware is learning to think on its feet, it is worth remembering what still stops an account takeover cold. A hardware security key like the YubiKey 5 NFC is one of them. Even if an attacker has your password, they cannot log in without the physical key in hand, and it takes about two minutes to set up on your main accounts. It is not glamorous and it will not fix everything, but for the price of a takeaway it closes the door that most attacks walk straight through. Check the price on Amazon.
Quick one before you go: if you could hand one part of your workflow to an AI agent tomorrow, what would you trust it with, and what would you never let it touch? Hit reply and tell me, I read every one.
Nath, Future Technology
Forwarded this? Get your own free briefing at futuretechnologyhq.com/newsletter
Some links may be affiliate links.