Envision claims first commercial 800V DC data center power system in Inner Mongolia

Tom's Hardware reported the announcement. The system has been in testing since August at the Galaxy campus. Envision says the site will scale to 3GW of capacity under current plans. It says the site will eventually "support up to one million AI accelerators."

The claim is the company's own. No independent measurement of the efficiency gain has been published. The press release quoted by Tom's Hardware gives no figure for how much power the conversion path saves against conventional designs. That absence is the first thing to note. A vendor announcing a first is not evidence of a first. The release offers no baseline, no test protocol and no third-party audit of the result.

What the 800V DC architecture actually changes

Data centers today step grid power down through several conversion stages before it reaches a server rack. Each stage loses energy. Each one also takes up floor space. Envision's Dubhe system converts directly from 10kV medium-voltage AC to an 800V DC bus. That cuts the number of stages and shrinks the conversion equipment, according to Tom's Hardware's account of the company's release.

The company lists three capabilities behind the design. A solid-state transformer holds stable conversion from 10kV to the 800V bus across operating conditions. Battery storage provides both second-scale power support and minute-scale backup. Millisecond-level coordination between the transformer and the storage absorbs GPU load swings locally. The battery element matters because AI training loads are spiky. Smoothing them at the rack rather than at the grid interface is the part of the pitch that differs from a plain voltage change.

Envision is not a neutral party here. It sells the controller, the storage and the energy management software. The Galaxy site is its own. The deployment proves the hardware runs in a commercial setting. It does not yet show what it costs or how it compares with the AC distribution most operators still buy.

The component squeeze reaches gamers first

The same demand driving 3GW campus plans is visible further down the market. Valve's October Steam Survey was posted without comment as usual and covered by Tom's Hardware on 3 November. It shows the share of systems with 32GB of RAM falling 5.16 percentage points in a single month. In the previous survey, 32GB had been the most popular configuration. The proportion of players with 1TB or more of total storage also fell by nearly 5%.

Tom's Hardware frames the shift as a symptom of the AI-driven components crunch. It notes that prebuilt gaming systems are now frequently fitted with 16GB as standard. That is an inference from the survey data rather than something Valve states. The survey is a voluntary sample of Steam users. A one-month swing of that size is unusual enough that some of it may reflect a change in who responded rather than a mass downgrade.

Only last month, 32GB was the most popular configuration among gamers according to the same survey.

Either way, the direction is consistent with what memory pricing has done through 2026. Tom's Hardware maintains a RAM price tracking page. It reports that the current market makes DIY upgrades expensive. The data center buildout and the desktop upgrade path are competing for the same DRAM supply.

Older hardware loses its software first

Support endings are the quieter side of the same story. Heise reported on 3 November that Meta will stop developing new system features for the Quest 2 and Quest Pro. Distribution of Horizon OS 2.13 is due to begin in January 2027 and finish by March. After that, new system updates, features and APIs go only to Quest 3, Quest 3S and the coming Meta VR Glasses.

Meta said in its developer blog, per heise, that the older headsets will get quarterly security updates and critical bug fixes only, through the end of 2027. Developers can submit new apps and games for Quest 2 and Quest Pro until December 2027. After that only updates to existing applications will appear in the Horizon Store.

Purchased apps stay playable and the store stays open on both devices, heise reports. Cross-play with Quest 3 and Quest 3S users continues, as does Quest Link for PC VR. Sales of the two headsets ended in 2024. The Quest 2, launched roughly six years ago, is estimated to have sold more than 20 million units. It remains the best-selling VR headset, and Meta says millions of people still use it.

Heise notes that developers now face a choice. They can maintain one backward-compatible build or ship separate versions for old and new hardware. Some will drop support. Where a key online service is switched off, a game can stop working entirely. Developers who cut off access to purchased content or important services must give users at least 180 days' notice.

Design tools strain under the same complexity

Semiconductor Engineering's 2 November roundtable on shift left puts a related problem in engineering terms. Frank Schirrmeister of Synopsys described the design flow as "a labyrinth." Packaging, thermal, electromagnetic and mechanical constraints all have to be modelled together. A packaging decision can be blocked for a year by a constraint nobody modelled early.

Ashish Darbari of Axiomise told the roundtable that bug escapes now pull architects, designers, verification leads and software teams into the same room. They ask why a defect got through, and each group holds a different idea of what the specification said. Jack Berg of Silvaco said verification has traditionally been the longest and most expensive part of a design. He said that "how to shift everything left is something we're still learning." The panel agreed that agentic AI may improve coordination across disciplines but cannot replace simulation and emulation.

That is the less visible cost of denser hardware. The validation work grows faster than the transistor budget shrinks.

Governance experiments run alongside

A paper posted to arXiv on 29 October and listed on 2 November documents a pilot expert workshop on "violet teaming." The method folds technical, policy and societal views into the evaluation of emerging technologies. The authors, Steph Batalis and four co-authors, ran a hypothetical biodefense-relevant AI scenario through three stages: capability development, misuse potential and implementation.

According to the abstract, the exercise surfaced non-obvious misuse pathways. It expanded the range of possible interventions beyond purely technical safeguards, while also highlighting operational and policy constraints. It is a nine-page report of one workshop, not an evaluation framework with results. The authors present it as lessons learned rather than a validated method.

Further from the data center, University of Michigan-led research examined the binary system 70 Ophiuchi AB, 16 light-years away. The work was published in The Astrophysical Journal and reported by ScienceDaily on 29 October. The team found no Jupiter-sized planets in the regions it searched, ruling out giant worlds there. That leaves open the possibility of Saturn-sized or smaller planets in the habitable zone. Yiting Li, a postdoctoral researcher at Michigan, said: "It's a mix of disappointment and excitement." The system has drawn planet claims since William Jacob's proposal in 1855.

Istanbul Technical University reported on 28 October that an international team had identified a previously unrecognized volcanic system in northwestern Iran. It is named the Azerbaijan Continental Magmatic Arc and was active roughly 101 to 97 million years ago. The work, published in Geoscience Frontiers, used zircon U-Pb geochronology, Hf isotope analysis and whole-rock geochemistry. Corresponding author Ali Mohammadi said the study shows the region's magmatic history "is more complex than previously recognized."

IEEE Spectrum reported on 28 October on Beta Technologies, one of the top three U.S. electric aircraft makers. The company is targeting regional cargo and medical transport before the passenger market. It is building motors, batteries and a charging network alongside its Alia aircraft. The company's bet is that the supporting infrastructure, not the airframe alone, decides whether electric flight works.

None of this settles whether Envision's 800V bus is the right answer for AI power delivery. It does show that the constraint has moved. The bottleneck is now the conversion and storage hardware around the accelerators, and the memory and software support cycles behind them, rather than the chips alone.