Oks argues that hyperscaler demand for HBM3E and high-density DDR5 is forcing Samsung, SK Hynix, and Micron to redirect fab capacity away from mobile LPDDR, with DRAM contract prices up ~30% in Q3 2025 and another double-digit jump baked into Q4. He frames this as a quiet but structural repricing of the budget Android tier, not a temporary supply blip, citing HBM eating 35-40% of advanced DRAM capacity versus under 10% in 2023.
The editorial extends Oks's thesis to the second-order effect most trade press has missed: Transsion, Tecno, Infinix and Shenzhen ODMs operate on 3-5% gross margins, so a memory BOM line jumping from $18 to $32 on a $150 device forces either price hikes, spec downgrades, or SKU elimination. This matters because the cheap-Android tier has been the internet on-ramp for roughly two billion people, and compute is now being rationed globally in favor of AI workloads.
A post by David Oks titled *AI is killing the cheap smartphone* hit the top of Hacker News with 319 points, arguing that the global memory shortage — driven by hyperscaler appetite for HBM3E and high-density DDR5 — is repricing the entire consumer electronics stack from the bottom up. The thesis is uncomfortably simple: when Samsung, SK Hynix, and Micron tilt their fabs toward AI memory, the wafers that used to become $40 LPDDR4X packages for a $180 Android phone don't exist anymore. The result is a quiet repricing of the budget tier that has, until now, been the workhorse of the global smartphone market.
The numbers behind the post are not subtle. DRAM contract prices rose roughly 30% in Q3 2025 and another double-digit jump is already baked into Q4 contracts, with NAND following on a similar curve. TrendForce and Counterpoint have both flagged HBM as eating 35-40% of advanced DRAM capacity at the leading-edge nodes, up from under 10% in 2023. SK Hynix is reportedly sold out of HBM3E through 2026. Samsung's DS division has been redirecting 1a/1b nm capacity away from mobile LPDDR. Micron's last earnings call put the bit growth allocation for AI data center memory at "materially above" what the company can supply.
What the HN thread surfaced — and what most of the trade press has missed — is the second-order effect on the OEMs that build at the bottom of the market. Transsion, Tecno, Infinix, the white-label ODMs in Shenzhen — these companies operate on 3-5% gross margins. When the memory BOM line jumps from $18 to $32 on a $150 device, there is no margin to absorb it. Either the price goes up, the spec goes down, or the SKU disappears.
The interesting part isn't that phones are getting more expensive. It's *which* phones, and what that signals about how compute is being rationed globally. For the last decade, the cheap-Android tier has been the on-ramp to the internet for roughly two billion people. That on-ramp is now being priced by the same forces that set the cost of an H100 cluster in Ashburn. The marginal gigabyte of DRAM is no longer allocated by consumer demand; it's allocated by whichever buyer can monetize a token fastest, and a hyperscaler training run beats a Lagos teenager's first smartphone every time.
This is the part the original post gestures at but doesn't finish: the mechanism isn't just "AI demand high, prices up." It's that memory has become the new bottleneck in the same way GPUs were the bottleneck in 2023. Nvidia's supply constraint moved upstream. HBM is built on the same DRAM lines as LPDDR and DDR5, with shared lithography, shared packaging capacity (especially CoWoS at TSMC), and shared 1a/1b nm DUV tooling. You cannot decouple the AI memory market from the consumer memory market on a 12-18 month horizon, because the fabs are physically the same.
The community reaction on HN was split in a telling way. The hardware engineers in the thread mostly agreed with the diagnosis but pushed back on the timeline — pointing out that new fab capacity from Samsung's P4 and Micron's Idaho expansion comes online in 2027, not 2026, and that HBM4 will only intensify the squeeze. The investors in the thread treated it as a Micron/Hynix long thesis; the developers treated it as a cost-of-living story; almost nobody treated it as what it actually is, which is a quiet rewrite of who gets to compute.
There's also a worth-noting comparison to the 2017-2018 NAND shortage, which was demand-driven (datacenter SSDs) but resolved within ~18 months as Chinese fabs (YMTC) ramped. This time the demand is structural — every major model training run wants more HBM, not less — and the supply response is gated by EUV tools that are themselves capacity-constrained at ASML. The Wright's-law curve that delivered cheaper memory every year for thirty years has, for the first time since the early 90s, inverted.
If you ship software, this is going to land on your roadmap in three places, probably without anyone naming it. First, cloud bills. AWS, GCP, and Azure all price memory-heavy instance types (r-class, m-class with large RAM) against the same DRAM supply curve. Spot prices for r7i and r7g instances have already crept up in 2025, and reserved-instance pricing for 2026 contracts is being negotiated against a tighter cost base. If your service is memory-bound — Redis, large-heap JVMs, in-memory analytics — model a 10-20% increase in per-GB cost into your 2026 capacity plan.
Second, SSDs and developer hardware. NAND tracks DRAM with a lag. The cheap 2TB NVMe drive that costs $90 today will be $130-150 by mid-2026 if current contract trends hold. Buying laptops, dev workstations, or homelab gear? Pull the purchase forward into Q1 2026 if you can; the curve doesn't bend back until late 2027 at the earliest.
Third, on-device AI strategies. Every "run the model locally" pitch — Apple Intelligence, Pixel's Gemini Nano, Qualcomm's on-device Llama — assumes memory is cheap enough to ship 8-16GB of LPDDR on a mid-range device. That assumption is now wrong at the budget tier. Expect mid-range Android phones to ship with less RAM for the next 18 months, which means the on-device LLM story fragments along price lines. If you're shipping a mobile app that depends on local inference, your addressable market just got smaller.
There's a quieter implication for hiring and geography too. The HN post's strongest claim — that the cheap smartphone is dying — translates directly into fewer first-time internet users in emerging markets over the next two years. If your product depends on next-billion-user growth, the unit economics of that audience just shifted.
The memory shortage is not a cyclical blip. It is the consumer-side bill for the AI capex cycle, and it will arrive on every BOM, every cloud invoice, and every laptop refresh quote between now and 2028. The fab capacity that would relieve it is two years out, and HBM4 will absorb most of it on arrival. The practitioner move is to stop treating memory as the cheap, abundant resource it was for thirty years, and start budgeting it like you budget GPUs: a scarce input with a queue.
The MacBook Pro on which I’m writing this piece needs memory that can keep up with a powerful processor running many programs at once: so it uses a standard called DDR, “double data rate,” which runs at a reasonably high voltage and offers high bandwidth. The processor on my iPhone is less powerful,
What was most surprising about all this to me was this line:> So modern DRAM manufacturing is an extraordinarily complex and expensive process. Building a single state-of-the-art DRAM fabrication facility, a “fab,” will cost you about $15 to $20 billion; acquiring all the necessary equipment, lik
It's impressive that somehow, as if by coincidence, we're seeing the biggest inflationary drivers for decades, perhaps for centuries, all happening simultaneously.The Iran war is spiking the price of oil and will likely cause shortages of pretty much everything if it isn't ended.The U
So the availability of cheap phones is going down because of the cost of RAM.What about the RAM consumption trend of the last 10 years? I think it is very feasible to produce phones with the same amount of RAM as was the norm 10 years ago. The only compromise would be using older algorithms and feat
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
The headline here under-serves the article in my opinion: this is a fascinating, deep explanation of how the memory market works and why increased demand for HBM (used by big GPU racks) hurts the availability of wafers for DDR and LPDDR (used by laptops and phones).