You know how every few months someone declares the AI industry is about to change forever? This week actually felt like one of those moments — and not for the usual hype reasons. Between Samsung posting an 18-fold profit jump on memory chips, Chinese AI models making Western pricing look almost comical, and Palantir's CEO telling everyone to stop handing their data to LLM companies, there's a lot to unpack.
The Chinese AI Pricing Reality Check
I've been running GPT-4o for most of my daily workflow — coding snippets, content drafts, the occasional brainstorming session. Then I came across a Dev.to post from someone who spent two weeks testing Chinese AI models, and honestly, the numbers made me double-take.
DeepSeek V4 Flash costs $0.25 per million output tokens. Qwen3-8B costs literally one cent. One cent per million tokens. For reference, that's roughly 750,000 words. You could process an entire novel for less than a gumball.
The comparison between four model families — DeepSeek, Qwen, Kimi, and GLM — paints a pretty clear picture:
- DeepSeek V4 Flash ($0.25/M) is the daily driver pick for most people on a budget. Solid code generation that holds up against models ten times the price. I threw some API integration work at it this week and honestly couldn't tell the difference from GPT-4o. The wallet sure can.
- Qwen offers the widest range — from the absurdly cheap Qwen3-8B at $0.01/M all the way up to Qwen3.5-397B at $2.34/M. If you need options, these guys have them.
- Kimi K2.5 ($3.00/M) punches above its weight on reasoning tasks, though you're paying flagship prices for it.
- GLM crushes Chinese-language work and starts at that same $0.01/M floor.
To be fair, there are gaps. DeepSeek's vision capabilities are weak — if you need image analysis, you're looking elsewhere. The English nuance on some Chinese-native models still doesn't quite match Claude or GPT-4o on creative writing. But for pure price-to-performance on the kind of work most developers actually do — coding, summarization, structured output — the gap is closing fast.
I switched my daily coding assistant to DeepSeek V4 Flash last week. For API integrations, debugging, short scripts, it's been indistinguishable. My monthly API bill dropped by about 80%.
Samsung's 18x Profit Jump — The Agentic AI Angle Nobody's Talking About
Samsung is expected to report an 18-fold jump in operating profit — 86 trillion won ($56 billion) for Q2 — driven entirely by AI demand for memory. DRAM prices jumped 44% quarter-over-quarter. NAND went up 53%. Samsung, SK Hynix, and Micron have all crossed $1 trillion market cap this year. That's not a typo.
The usual narrative pins this on HBM for training clusters. But analysts are now pointing at something else: agentic AI. Systems that don't just generate text but actually perform multi-step tasks — booking flights, managing workflows, running code — require way more conventional DRAM and NAND during inference. We're past the point where AI is just about training giant models once. The inference phase is where the real memory hunger kicks in.
There's a darker side though. JPMorgan flagged that AI memory now eats up 52% of cloud providers' capex, projected to hit 70% next year. That level of concentration makes even bulls nervous. If AI service revenue doesn't materialize fast enough to justify the spending, the boom could cool off quicker than people expect. Samsung's mobile business is already feeling the pinch — rising memory prices are squeezing phone margins despite recent price hikes.
Palantir's CEO Drops a Truth Bomb
Alex Karp published a nine-point manifesto that boils down to a simple message: stop handing your proprietary data to LLM API companies. His argument — the companies selling tokens refuse to eat their own dog food, and there's a reason for that. It's blunt, but not entirely wrong.
A startup I've been talking to spent months building on top of an API-based LLM, only to realize they had zero control over the data pipeline. They ended up fine-tuning a smaller open model for their specific use case and got noticeably better results. From my perspective, the "build on open, use APIs for exploration" hybrid approach is where most teams should be looking right now.
Cloudflare Steps Into the AI Search Mess
Cloudflare announced two initiatives around AI search this week. The core problem is brutally simple: AI summaries destroy click-through rates. A Pew study showed that when Google serves an AI summary, users click a traditional link just 8% of the time — about half of what it is without the summary. For creators who rely on traffic for revenue, that's existential.
Cloudflare's play is to give site owners better bot control while trying to rebuild the economic model — making AI search "smarter" by using the signals they already see across their network. Whether it works is another question, but at least someone's trying to solve the creator visibility problem instead of just playing whack-a-mole with crawlers.
The Absurd AI PC and the Security Reality Check
TechRadar spotted a mini PC packing an AMD Ryzen AI Max 395, 128GB of RAM, 126 TOPS of NPU performance, and — I'm not making this up — a vegan leather handle. It's ridiculous and I kind of want one. For anyone running local models, even medium-sized ones, that kind of memory headroom is genuinely useful. You can run decent-sized LLMs on the NPU without touching the GPU. Not a workstation replacement, but a compact AI sandbox? Tempting.
Meanwhile, Ars Technica published research showing that AI-powered browsers can be tricked into ignoring their guardrails. Every new capability opens a new attack surface, and we're still very much in the early days of figuring out what AI agent security looks like. Worth keeping in the back of your mind if you're building anything agentic.
Quick Hits
- India's Mohandas Pai: advised the country not to chase the costly LLM race and focus on practical AI models instead.
- Base44: built its own LLM specifically to combat what the CEO calls "AI-slop design" — an interesting case for companies that want control over output quality.
- Data protection in Europe: new study shows GDPR-related rules are slowing LLM rollout across the continent, creating a growing regulatory gap with the US and Asia.
- E-commerce startup Lantern: pivoted to GEO (generative engine optimization) and agentic marketing tools — another sign that the SEO-to-GEO shift is real.
Value is shifting in three directions at once. Model inference is getting absurdly cheap thanks to Chinese competition. Memory and infrastructure are getting absurdly expensive thanks to agentic AI demand. And the data ownership debate is heating up in ways that will shape the next phase of the industry.
If you're building with AI today, experiment with the cheaper models. You might be surprised at how little you actually lose by switching. And keep an eye on your memory costs — they're going up whether you're buying chips or renting cloud instances.
If you've tried DeepSeek or Qwen recently, I'd love to hear how it compares to your usual stack. The pricing gap is one thing — the actual experience is another. Drop a comment and let me know what's working for you.
Built something interesting this week? Check out PayCalc for quick salary and compensation calculations.
Top comments (0)