[SMM Computing Power Daily] DDR5 Price Spread Halved, PRO6000 Quoted Price Difference at 2.4x

SMM, September 22:

[DDR5: Channel sellers unilaterally cut prices by 5.9%, bid-ask spread halved to 1,157 yuan]

SMM's weekly monitoring of first-hand information on DDR5 server memory in China's channel market shows that in the most recent week (September 14-20), sellers visibly turned anxious: the shipment heat index rose from a base of 100 to 157.3, up 57.3%, while purchase heat only rose to 102.8; buyer willingness intensity fell from 48.8% to 36.3%, down 12.5 percentage points, the lowest range since mid-July. On pricing, the median shipment asking price fell from 19,165 yuan to 18,036 yuan, down 5.9%, while the median purchase bid price edged up from 16,732 yuan to 16,879 yuan, narrowing the bid-ask spread from 2,433 yuan to 1,157 yuan—a convergence driven entirely by sellers, not by buyers bidding up. Channel data as of September 18 shows Samsung 64G 5600/6400 tax-inclusive shipment quotes at 18,100-18,350 yuan, with actual purchase orders for the same specs at 18,000-18,200 yuan, a high-low gap of just 100-300 yuan. Over the same period, the contract side at original manufacturers moved in the opposite direction: CFM Flash Market reported on September 8 that DDR5 RDIMM 64GB contract prices were $1,630, up 3% MoM, and TrendForce projected 3Q26 server DRAM contract prices to rise 13%-18% QoQ. SMM believes channel inventory realization pressure has overtaken shortage premiums, and DDR5 pricing power is shifting from unilateral seller dominance to proactive seller concessions to narrow spreads.

[PRO6000: One name, four specs—dispersion under 4% once disaggregated]

SMM's item-by-item review of first-hand quotes for the PRO6000 series in China's channel market from September 14-22 confirms that this series was the hottest-quoted GPU category this week, ahead of the 5090 and H200; but under the same "PRO6000" name, the 96G server version was quoted at 160,000-166,000 yuan/piece, the 96G workstation version at 134,000-137,000 yuan/unit, the 84G PRO6000D at 68,300-70,000 yuan/piece, and the Max-Q at 66,500-111,000 yuan/piece—a 2.43x gap between the highest and lowest. The most easily confused are the 96G server version and the 84G PRO6000D—only 12GB of VRAM apart but a 2.37x price difference, with both sharing the same name in channel parlance. Once disaggregated by spec, the three mainstream specs showed dispersion of 3.8%, 2.2%, and 2.5% respectively, and bid and ask prices for the 96G server version already overlapped; only about 12.8% of channel quotes provided verifiable figures. SMM believes the issue is not chaotic quoting or market fluctuations, but a dual absence of a naming system and a public price anchor—the series has no official direct-supply channel in China, so price discovery relies solely on verbal channel quotes.

[4090: Sichuan 8-card full server annual rental at 7,300 yuan, bandwidth listed separately]

SMM has learned that a data center in Sichuan publicly listed five 8-card RTX 4090 (24G) full servers for rent at 7,300 yuan/unit·month (on an 8-card full-server basis), equivalent to approximately 912.5 yuan/card·month, with total monthly rental scale of 36,500 yuan for the five units; bandwidth is charged separately at 10 yuan/M and not included in the machine-hour price. All four elements (price, lease term, unit count, data center location) are complete, and the listing is taken at face value as publicly posted. SMM believes the rent itself falls within the existing 4090 range, and what truly affects future price comparisons is the separation of bandwidth from the machine-hour price—once machine-hour price and bandwidth are quoted separately, the two quotes must first be aligned on whether bandwidth is included before they are comparable. (This item was published as [SMM Computing Power Flash] on September 22.)

[Other market developments · Full servers and supply-demand: H200 supply signals surface, no prices quoted]

SMM has learned that two H200 leads appeared on the supply side but neither included pricing: one is a white-label full-server factory processing order, with 1T/2T configurations available, cash on delivery, daily pickup of 5-10 units, 512 units deliverable before year-end with warranty, requiring a framework agreement; the other is a 64-unit long-term rental listing, 5-year closed term, one-month deposit and one-month advance payment, IB networking, data center located in Xinjiang, price not specified. Both are single-channel public group messages without price benchmarks and were not included in price statistics. On the memory side, buyers continued to post purchase requests for 2,000 units each of Samsung and Hynix 64G 5600, DC25 vintage, stopping once the target is reached. SMM believes the H200 large orders are becoming clearer on volume and delivery terms, but sellers have yet to reveal pricing, consistent with the recent "surface without quoting" bargaining posture in the high-end full-server segment.

[Policy updates: Five ministries set hard utilization threshold; Jiangsu and Hubei expand token vouchers]

SMM has learned that the Ministry of Industry and Information Technology and four other ministries jointly launched the 2026 national green computing infrastructure recommendation program, with hard thresholds including at least one year of continuous stable operation, cabinet utilization rate of no less than 60%, and rack scale of no fewer than 3,000 standard racks (excluding public institutions, financial computing, and intelligent computing facilities). The evaluation framework raises the weighting of carbon efficiency, green electricity share, and utilization rate. Submissions must be completed by November 16, with dynamic list management—facilities failing key indicators will be removed. On the local front, Jiangsu issued a token economy action plan with 26 measures across eight sectors, including building 2,000 high-quality datasets, adding no fewer than 50 newly registered large models annually, exploring token metering systems and circulation trading platforms, and rolling out token vouchers province-wide; Hubei added model and token service models to the scope of computing power voucher subsidies. On the China-US front, economic and trade consultations in New York on September 20 lasted about eight hours, with both sides establishing a "US-China AI Dialogue" mechanism and a dedicated national security-level AI incident notification hotline. The US side made clear that chip export controls were not on the agenda for this round, and the next meeting will be held in Shenzhen in late November. SMM believes the hard utilization threshold and dynamic elimination mechanism are efficiency metrics designed to force out low-efficiency capacity, advancing in tandem with the expansion of token vouchers to the full "model + token" chain on both the supply and demand sides.

SMM analysis: Four main threads today. First, the memory channel market saw the first instance of unilateral seller price concessions—shipment heat rose 57.3% while purchases rose only 2.8%, median asking prices fell 5.9%, and the spread halved through proactive seller moves, running counter to the rising contract side; the divergence between spot and contract prices warrants continued monitoring. Second, PRO6000's apparent price spread stems from the absence of a naming system and price anchor, not market volatility—once the four specs are disaggregated, dispersion is below 4% for all, and bid-ask prices for the server version already overlap, indicating channel consensus exists while a public price discovery mechanism is absent. Third, 4090 rents are stable but the quoting basis is shifting—the Sichuan 7,300 yuan listing falls within the existing range, but the separate listing of bandwidth changes quote comparability (this item has been published as a flash update and incorporated into the comparison checklist: whether bandwidth is included, full server or single card, lease term and payment, IP and electricity/warranty attribution—quotes are excluded from price statistics until these four items are aligned); meanwhile, a retrospective verification confirmed that the 4090 buyer share spike from July 28 to August 3 was driven by concentrated posting from recyclers and card dismantlers (during the spike window, 79 dismantling-related posts versus 16 rental-related posts, with the top 5 accounts accounting for 40%; over the same period, SMM monitored 4090 PCB purchase prices above 18,000 yuan), representing noise in the data rather than a pickup in end-use demand—heat indicators must first strip out recycling-related posts. Fourth, on the policy front, efficiency thresholds and demand subsidies are being tightened simultaneously, while the China-US AI dialogue mechanism has been established and chip export controls were not on the agenda, meaning geopolitical uncertainty in the advanced computing supply chain is unlikely to dissipate in the near term.