Should we buy more RTX 5090 nodes now? A first-principles price analysis#

Status: Analysis Last verified: 2026-09-27, against the primary sources marked "read directly" below and a SemiAnalysis pass recorded in the claims ledger Canonical for: where the cost of our node is heading over the next 12, 24 and 36 months, and why

The question. Owning compute is settled (DECISIONS 2026-09-26). This page asks whether today's prices are a peak we'd regret buying into, or a floor that only rises. It answers from how each part's price is set, applied to our actual node.

Partial observability. Every number here is an estimate from a limited view. My web access is narrower than a search engine's, and some trackers block it. Inputs are labelled:

Where they conflict, the reasoning about how the price is set outranks any single forecast.

Claim IDs in square brackets, such as [TF-0730-2], point to rows in the claims ledger. Each row records the source, what corroborates it, our verdict, and what would prove it wrong. If a claim's verdict changes, re-check every conclusion here that cites it.

1. What's in a node (our data)#

A PC2-like node costs about RM36.2k, including its OS drive.

About 70% of the node is priced by memory: directly through RAM, and indirectly through the GPU's 32 GB of GDDR7.

2. DRAM: up through 2027 (the most solid part)#

How the price is set.

• Supply:
  • Three makers (Samsung, SK hynix, Micron) plus CXMT, China's DRAM maker.
  • Fab space is fixed for about two years; a new fab takes that long to build and equip.
  • HBM, the stacked memory on AI accelerators, uses roughly 3× the wafer per bit of ordinary DRAM, and HBM4 and HBM4E make that worse [SA-silicon-5]. It earns makers more. So capacity moves to HBM and server memory first, and consumer DDR5 and GDDR7 get what's left.
  • Shrinking the process no longer adds bits cheaply: DRAM density grew only about 2× in a decade [SA-memmania-1]. New bits need new wafers, meaning new fabs.
• Demand:
  • AI server buildout, with memory per server rising.
  • Consumer PCs and phones, which are price-sensitive.

Evidence.

What could break it before 2028.

1. An AI spending correction. The strain is visible: Alphabet reported negative free cash flow last quarter, its first since its 2004 IPO; Meta was about zero; spending is increasingly debt-funded; and investors pushed back after July earnings. That is a late-cycle sign, but guidance is still going up. If spending were cut, memory would likely ease about two quarters later, once wafers move from HBM back to ordinary DRAM.
2. CXMT: relief at the margin, not a cycle-breaker (corrected 2026-09-27).
   • Growth: about 350k wafers a month at end 2026, 420k in 2027 and 500k in 2028 [SA-cxmt-1]. Bit share rises from about 9% to about 12% by 2027 [SA-cxmt-2].
   • Limits: its cost per bit is 30%+ above the incumbents on an older process [SA-cxmt-3].
   • It is fungible. Every Chinese maker that switches to CXMT frees incumbent supply for everyone else.
   • But the scale is small. About 3 points of share over two years eases the shortage; it doesn't break it.
   • An earlier version of this page called CXMT "the most underrated reason consumer DDR5 could ease early". SemiAnalysis's capacity and cost figures don't support that.

A plateau rather than an ever-steeper climb. Profit on ordinary DDR has reportedly caught up with HBM contract levels [SA-memmania-4, pending]. If so, makers have no reason to shift even more wafers to HBM. DDR supply stops getting worse, but it doesn't get better until the 2028 fabs arrive.

RAM for a new node:

• 12 months: +20% to +50% (+RM1.0–2.5k). The rise slows as PC demand shrinks and the HBM shift plateaus.
• 24 months: −20% to +30% on today's price. Relief starts in 2028, but prices stay high through it [TF-0730-2, SA-dwarkesh-1].

3. NAND (SSDs): the exception#

TrendForce (same release, read directly) expects NAND to swing to surplus in the second half of 2027 [TF-0730-4]. Three things drive that:

• adding flash layers raises bit output cheaply;
• new fabs are ramping;
• consumer demand is weak.

SSDs:

• 12 months: 0% to +10% (+RM0–0.4k).
• 24 months: −10% to −30%.

4. The GPU: the part the reports got wrong#

How the price is set. The RTX 5090 and the RTX PRO 6000 are cut from the same die, GB202 (read directly, Tom's Hardware).

Nvidia earns about 4× more per die, and more per GB of the scarce memory, on the PRO card. Gaming is about 9% of Nvidia's revenue. While PRO cards sell, every die sent to a 5090 costs Nvidia money. The rational move is to starve the 5090, and that is what's reported: gaming supply cut by 20–40% in 2026 (several outlets, reported).

The assumption behind the reports' price decline has broken.

• Report B forecast the 5090 falling 18–29% over 12 months, assuming the next generation (RTX 60, "Rubin") launched in late 2027.
• On 21 Sep 2026 the leaker kopite7kimi, who has a strong track record, said Nvidia had postponed it to
  2028. The stated reason is memory cost. This is reported, and a rumour.
• The RTX 50 Super refresh is also stalled on the cost of 3 GB GDDR7 chips (reported).
• Both fit the mechanism above: Nvidia won't launch consumer products into a memory shortage.

There's a physical reason for the delay, not just a leak.

• TSMC's N3 process is the binding constraint on AI chips [SA-silicon-1]. AI takes about 86% of N3 in 2027 [SA-silicon-2], and Nvidia alone holds 70%+ [SA-dwarkesh-2].
• A Rubin-generation consumer die would need N3, and there are no N3 wafers for it in 2027. A 2028 launch lines up with Nvidia's data-centre parts moving on from N3.

Wafers aren't what limits the 5090. GB202 is on TSMC's 4N family. As Nvidia's data-centre volume moves from 4NP (Blackwell) to N3P (Rubin) [SA-silicon-4], 4N pressure eases. What limits the 5090 is GDDR7 and Nvidia's choice to send GB202 dies to the PRO card.

Demand.

• 32 GB at a consumer price has no substitute.
• AI buyers, us included, absorb stock as it lands.
• There is grey-market demand from China.
• Better open video models need more VRAM, not less.

Local observations.

What would push the GPU down.

• RTX 60 arriving (now 2028).
• Memory easing (2028).
• An AI spending correction that crashes demand for the PRO cards.
• Nvidia choosing to flood gaming, which it has no incentive to do.

None of these clearly lands within the next 12 months.

GPU:

• 12 months: flat to +20% (+RM0–4.2k). The only path down in that window is an AI spending correction.
• 24 months (September 2028): −25% to +10%. RTX 60 may be out by then, but it launches into memory that is still expensive. The 5090 falls somewhat, not by half.

5. Adding it up#

At 24 months, in RM: GPU −5.2k to +2.1k; RAM −1.0k to +1.5k; SSD −1.2k to −0.4k. Node total −7.8k to +3.2k on RM36.2k.

Waiting two years probably saves about 5% (anywhere from −21% to +9%) and costs two years of the node's work. The first real discount is about three years out, not two. An earlier Gemini-based view said 20–40% cheaper by 2028. With SemiAnalysis's supply evidence, relief comes later and is shallower.

The case where waiting would have won: an AI spending correction in 2027. The stress signals are real:

• Alphabet reported negative free cash flow;
• Meta was about zero;
• the case that returns on the infrastructure turn out middling [SA-dwarkesh-6].

But guidance is still rising. There's no honest number for its odds; it is a real minority risk, not the base case.

The early warning is the rental price of old GPUs. A glut shows first as three-year-old GPUs renting for less. H100 1-year contracts rose about 40%, from $1.70 (October 2025) to $2.35 (March 2026), with on-demand capacity sold out [SA-rental-1, SA-rental-2]. The pattern that would change this call is the H100 contract index falling below $1.70 while hyperscalers still guide capex up [SA-rental-5]. Memory and GPU prices would likely follow about two quarters later. The latest reading we can see is from April 2026 (index note).

6. What this says#

1. Over the next 12 months, prices are more likely to rise than fall.
   • Memory rises almost regardless, and it is about 70% of the node through RAM and GDDR7.
   • The one force that would have pulled the GPU down in 2027, the next generation, has no N3 capacity until 2028.
2. Even two years out, the node is only about 5% cheaper at the centre, and could cost more. The first real discount is about three years out. Any node we'll put to work in the next two to three years is cheapest bought now. This supports the founder's decision to bulk buy now (DECISIONS 2026-09-26, update 2026-09-27).
3. Availability is a risk alongside price. MSI's store is sold out, and IdealTech kept last week's rise. If prices go parabolic, the constraint may be getting cards at all.
4. How many to buy is the founder's call. The price data says when; it can't say how many nodes the demand lever fills.

Correction on record#

An earlier answer in this session gave "+0–15% over 12 months" without computing it. It was skimmed from the Gemini reports, one of which rests on a made-up parts list. The range above is built from our node's parts. It lands near the same figure by luck, and it reverses the direction of the GPU piece. The Gemini-based synthesis.md now points here instead of carrying those figures.

Revised 2026-09-27 after a SemiAnalysis pass:

• the CXMT view was corrected;
• the N3 reason for the RTX 60 delay was added;
• 24- and 36-month rows were added, which make relief later and shallower than the Gemini view.

Sources#

• Claims ledger and source notes: every [ID] cited above
• SemiAnalysis: Memory Mania · The Great AI Silicon Shortage · The Great GPU Shortage: rental capacity · H100 index · Dwarkesh: Dylan Patel
• TrendForce: diverging memory outlook for 2027 (30 Jul 2026) (read directly)
• Micron Q3 FY26 8-K press release · Micron Q3 call transcript · Q4 due 30 Sep
• Tom's Hardware: RTX PRO 6000 shares GB202
• Tom's Hardware: RTX 60 timing · wccftech: RTX 60 in 2028
• PC Gamer: gaming GPU cuts up to 40% · TechTimes: supply cut 20%
• CNBC: capex scrutiny after Alphabet (28 Jul 2026) · TMT Finance: 2026 capex
• Gartner: memory costs cut PC and phone shipments · IDC: smartphones −16.7%
• TechPowerUp: CXMT near Micron capacity · SemiAnalysis on CXMT
• Tom's Hardware RAM price index
• MSI Malaysia store (read directly) · IdealTech RTX 5090 PC (read directly)