{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiegbrl4yt44bm5nlfb6v25e5pxdecz3ge5bkuqs4yxgqj7wl6tzsy",
    "uri": "at://did:plc:wnd7xrumusq5uayjfi2pgfno/app.bsky.feed.post/3me6zndk4u2t2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreigny2z5hopodzxnvwuw7ecsjmty3mngcl6cr7bptadppofz5y6obm"
    },
    "mimeType": "binary/octet-stream",
    "size": 252375
  },
  "description": "TL;DR\n\n * Dell and HP Shift DRAM Sourcing to Chinese Supplier CXMT Amid Global Memory Crunch\n * Envision Completes 12.8 GWh Energy Storage Cluster in Inner Mongolia, Grid-Connected in December 2025\n * SpaceX's Starlink Dominates Satellite Internet with 97% Speedtest Share and 9M+ Users\n\n\n📊 Dell & HP Qualify CXMT DRAM, Acer & Asus Test; Global DRAM Prices Spike 90% as Chinese Fabs Expand\n\nDell & HP just qualified CXMT’s DDR5/HBM2 DRAM—10-15% cheaper, 2-3x Chinese fab expansion underway. Global D",
  "path": "/2026-02-06-157221028535559630404442054726439582650/",
  "publishedAt": "2026-02-06T13:38:11.000Z",
  "site": "https://espresso.cafecito.tech",
  "textContent": "### TL;DR\n\n  * Dell and HP Shift DRAM Sourcing to Chinese Supplier CXMT Amid Global Memory Crunch\n  * Envision Completes 12.8 GWh Energy Storage Cluster in Inner Mongolia, Grid-Connected in December 2025\n  * SpaceX's Starlink Dominates Satellite Internet with 97% Speedtest Share and 9M+ Users\n\n\n\n* * *\n\n## 📊 Dell & HP Qualify CXMT DRAM, Acer & Asus Test; Global DRAM Prices Spike 90% as Chinese Fabs Expand\n\n> Dell & HP just qualified CXMT’s DDR5/HBM2 DRAM—10-15% cheaper, 2-3x Chinese fab expansion underway. Global DRAM prices spiked 90% QoQ; now OEMs diversify from Micron/Samsung. Can China’s memory surge stabilize AI server costs without risking supply chain resilience?\n\nDell and HP quietly added CXMT DDR5 modules to their approved-vendor lists on 5 Feb 2026, a move that instantly gives the Chinese supplier a double-digit share of two of the world’s largest PC memory budgets. The trigger is arithmetic: DDR5 contract prices jumped 90 % quarter-over-quarter in Q4 2025, and every 8 GB module now erases roughly 5 % of the OEM’s margin on an entry-level laptop. CXMT’s parts are quoted 10–15 % below Micron, Samsung and SK hynix, so the switch recovers half of that lost margin without raising sticker prices.\n\n### How Tight Is the DRAM Supply?\n\nAll three incumbent vendors are running fabs at 100 % utilization and still can’t meet demand. AI accelerators are soaking up every extra wafer of HBM2/3, forcing PC-grade DRAM lots onto the spot market where prices tripled in six months. CXMT, meanwhile, brought a new Shanghai fab on-line in late 2025 and doubled clean-room floor space in Wuhan; the company is on track to ship >50 % of its output as DDR5 by 2027. That capacity delta is the only visible buffer between OEM production schedules and a prolonged shortage.\n\n### What Technical Hurdles Remain?\n\nCXMT modules have passed Jedec-spec electrical tests, but field-failure databases are thin. Dell and HP are mitigating risk with 12-month burn-in clauses and dual-source contracts that keep Korean DRAM on critical SKUs. Acer and Asus, still in qualification, are sampling 4 Gb and 8 Gb DDR4/5 devices across 25 notebook boards; if error rates stay below 100 FIT at 85 °C, mass qualification will close before Q3.\n\n### Could Sanctions Upend the Deal?\n\nU.S. export rules already block Chinese fabs from 18 nm-class DRAM tools, yet CXMT obtained older DUV lines through domestic equipment vendors. A future sanctions package could freeze CXMT out of U.S.-designed controllers or EDA software, but not out of memory module sales already inside U.S. PCs. OEM lawyers have inserted force-majeure clauses that let them pivot back to Korean supply within one quarter if policy shifts.\n\n### Will Consumers Notice?\n\nNot on the spec sheet. Jedec timings, voltage and pin-out match legacy parts, and firmware already trains to CXMT SPD tables. The first Dell Inspiron and HP Pavilion units with Chinese DRAM ship in Q4 2026; if reliability mirrors early lab data, CXMT could command 30 % of global PC DRAM demand by 2027 and shave 15 % off AI-server memory costs once HBM2 ramps. For buyers, the only visible change is a price tag that stops climbing.\n\n* * *\n\n## ⚡ Envision Energy Deploys 12.8GWh LFP Storage, Enables Exascale HPC Power Stability\n\n> Envision Energy just launched the world’s largest single-phase LFP battery storage station — 12.8 GWh in Inner Mongolia, fully grid-connected and proven with 72h continuous discharge. This isn’t just energy storage — it’s the backbone for exascale HPC, AI clusters, and renewable stability. Can HPC data centers be the next major off-takers of grid-forming storage?\n\nInner Mongolia’s new Jingyi Chagan Hada battery park is not a pilot—it is a production-scale grid node.\n12.8 GWh of LiFePO₄ (LFP) containers, tied to the 800 kV trunk line in December 2025, passed three 100 % depth cycles plus a 72-hour continuous discharge at name-plate power. That 72-hour window is the new long-duration benchmark Chinese grid codes will require for every storage asset >100 MWh from 2026 onward.\n\n### Why LFP, and why now?\n\nLFP already dominates 98 % of 2025 Chinese storage additions because cycle life >10 000 and volumetric density 434 Wh L⁻¹ erase the nickel-cobalt price delta. Envision’s 530 Ah prismatic cells—stacked two thousand 10 MWh containers—deliver the same energy as 1.2 million urban households, but release it in sub-second response. The chemistry’s flat voltage curve keeps inverter efficiency >94 % round-trip, a figure previously reserved for pumped hydro.\n\n### What does “grid-forming” storage mean for compute?\n\nGrid-forming inverters source reactive power and hold frequency without external reference. For HPC campuses that means a 50 MW AI cluster can ride through a 500 ms transmission fault without spinning diesels. Inner Mongolia hosts 24 GW of wind; curtailment fell 12 % in 2025 simulations once the 12.8 GWh buffer was added. Expect colocation contracts offering <5 ms deviation and 99.999 % power quality—specs that match the U.S. exascale sites now under construction.\n\n### Where is the next 30 GWh?\n\nPipeline data show 20 GWh already permitted across Xilingol, Ulanqab and Alxa leagues. Envision’s domestic cell line runs at 60 GWh yr⁻1; containerized replication cuts CAPEX to $280 kWh⁻1, 30 % below European tenders. Policy tailwinds: NDRC mandates 180 GW nationwide by 2027 and spot-market participation for storage >50 MWh, turning arbitrage into a regulated revenue stream.\n\n### Bottom line\n\nA single province now operates more dispatchable electron capacity than the entire German battery fleet. For operators planning 2027-era exascale pods, the takeaway is concrete: firm, long-duration power is no longer a constraint—it is a commodity, traded by the millisecond, 12.8 GWh at a time.\n\n* * *\n\n## 🚀 Starlink Dominates Satellite Internet, Enables Edge HPC; 97.1% Market Share, 117 Mbps Speeds, 45ms Latency Enable Remote Supercomputing\n\n> Starlink now dominates global satellite internet with 97.1% market share, delivering 117 Mbps downlink & 45ms latency—enabling remote HPC job submission from rural labs. With 9,600+ satellites and BEAD funding, it’s becoming the backbone for edge-HPC clusters. Can satellite backhaul replace fiber for exascale data ingest?\n\nOokla’s Q3-2025 data set shows 97.1 % of all satellite-based Speedtest samples worldwide ran over Starlink; Viasat and HughesNet split the scraps at 1.7 % and 1.0 %. The constellation’s 8 300 operational LEO satellites deliver a median 117 Mbps down / 15 Mbps up at 45 ms latency in the U.S.—figures that beat geostationary rivals by > 50 Mbps and 600 ms respectively. With > 9 million active terminals, the network has become the de-facto reference for “satellite internet” in consumer-grade measurement tools.\n\n### Why does the HPC community care about a consumer ISP?\n\nRemote instruments, edge clusters and disaster-response labs often sit beyond fiber. A 117 Mbps LEO pipe with 45 ms RTT lets scientists ssh into supercomputers, push 50 GB checkpoint files or stream visualization frames without the 700 ms geostationary penalty. Starlink’s $50–120 retail tiers and $1 terminal-rental program drop the CapEx for a 10-node ARM cluster below the cost of a single month of leased fiber in many rural counties.\n\n### Where are the remaining bottlenecks?\n\nUrban canyons still see only 17–36 visible satellites, triggering 2–4 % packet loss during rush hour. The 60–140 W power draw of Gen-3 user terminals forces off-grid sites to budget 3.5 kWh/day—non-trivial for solar-battery rigs. Most importantly, only 44.7 % of U.S. tests hit the FCC’s 100/20 Mbps BEAD threshold; the rest fall short on upload, a problem for instrument data ingress.\n\n### What happens when the constellation doubles?\n\nSpaceX added ~7 500 birds in 2025-26 and targets > 10 000 operational by late 2026. Simulations show that 30+ passes per hour over CONUS will cut median latency to ≤ 40 ms and raise the BEAD-compliance share above 55 %. If the trend holds, a 200-node micro-datacenter on a mountaintop observatory could see sustained 1 Gb/s aggregated downlink—enough to feed a 10 PB/year sky survey directly into a national supercomputing facility.\n\n### Could regulators still clip Starlink’s wings?\n\nLight-pollution petitions and spectrum-sharing disputes loom, but the immediate risk is competitive: Amazon’s Project Kuiper and OneWeb’s Gen-2 plans promise sub-40 ms latency at comparable price points within 24 months. For HPC planners, the safest hedge is multi-orbit redundancy—pair Starlink terminals with terrestrial 5G or future Kuiper links and automate failover at the job-scheduler level.\n\n* * *\n\n### In Other News\n\n  * China implements national fixed-cost pricing for grid-side energy storage\n  * Raspberry Pi 4 Model B Rev 1.5 adopts dual DRAM modules to mitigate supply chain constraints\n  * Anker Solix Launches $3,799 Whole-Home Backup System for Residential Energy\n\n",
  "title": "Dell/HP Adopt Chinese DRAM, Envision Powers HPC with 12.8GWh Battery, Starlink Enables Satellite HPC",
  "updatedAt": "2026-02-06T13:38:11.000Z"
}