Cover Story · 01

Nvidia's shipping Rubin is slower on memory and NVLink than previewed — compute unchanged

ServeTheHome reported November 1 that Nvidia has quietly locked down lower shipping specifications for its Rubin GPU, just as the first Vera Rubin NVL72 systems reach general availability at hyperscalers. The revisions, posted to Nvidia's architecture datasheet in September and spotted by IT Home, cut HBM4 memory bandwidth from a preliminary 22 TB/s to 19.2 TB/s per GPU and NVLink scale-up bandwidth from 3.6 TB/s to 3.0 TB/s, while scale-out networking rose from 0.4 TB/s to 0.45 TB/s. Compute targets are untouched at 50 PFLOPS of NVFP4 inference and 35 PFLOPS of training, and memory capacity holds at 288 GB of HBM4 per GPU.

The rack-scale numbers move in lockstep: the Vera Rubin NVL72 drops from 1,580 TB/s to 1,400 TB/s of aggregate HBM4 bandwidth and from 260 TB/s to 216 TB/s of NVLink. The smaller HGX Rubin NVL8 takes the biggest hit — capacity down from 2.3 TB to 2.0 TB and bandwidth down 26%, from 176 TB/s to 130 TB/s, implying roughly 7 of each GPU's 8 HBM4 stacks are populated. Nvidia has said the earlier figures were preliminary targets and the new ones reflect qualified volume production; it has not said why the numbers moved. Even trimmed, Rubin's 9.375 Gbps/pin still sits above the HBM4 spec's 8 Gbps/pin and well above AMD's reported 7.6 Gbps/pin on MI455X — but buyers planning clusters around the original figures now have less memory and interconnect headroom than promised.

Source: ServeTheHome

02

Altman flies to Seoul as the memory crunch deepens

OpenAI CEO Sam Altman met about 50 Samsung Group executives in Yongin, south of Seoul, on October 31 to deepen cooperation on memory chips and AI infrastructure, Samsung said — a trip that also included a lunch meeting with SK Group chairman Chey Tae-won the same day. The talks build on the letter of intent Samsung and OpenAI signed last October covering memory supply for OpenAI's Stargate project, with Samsung SDS collaborating on AI data-center construction and enterprise AI services and Samsung C&T and Samsung Heavy Industries jointly developing floating offshore data centers. No new supply deal was announced, but the choreography is the message: with memory accounting for roughly a third of data-center spending and supplies tight, the model lab is negotiating directly with the memory makers, not just the GPU vendor.

Source: Aju Press · Hindu Business Line

03

The world's biggest data centers, ranked: Colossus 2 on top, Rainier biggest by power

Epoch AI's October ranking of the world's largest AI data centers, published October 31, puts xAI's Colossus 2 in Memphis first by compute at an estimated 1.1 million H100-equivalents, drawing about 946 MW and costing an estimated $35.8 billion so far on Nvidia B200 and B300 chips — with headroom toward 1.8 million H100e. Amazon's Anthropic-Amazon Rainier campus in New Carlisle, Indiana is second by compute but first by power draw at over 1,000 MW, running Amazon's own Trainium2 chips across a planned 30 buildings costing upward of $35 billion. Microsoft's Fairwater Atlanta (769,000 H100e, 636 MW, ~$24.1 billion) and OpenAI's Stargate Abilene (509,000 H100e, 421 MW, ~$15.9 billion) round out a top tier where the 10 largest sites alone represent well over $200 billion in capital spending. All figures are Epoch AI estimates built from satellite imagery, permit filings, and public disclosures, not official specifications — but the shape is clear: the frontier is now a handful of gigawatt-scale campuses, and two of the top 10 sit in Malaysia, not the US.

Source: OfficeChai / Epoch AI

04

Samsung's HBM share nears 40% as HBM4 takes over its memory mix

Industry estimates cited in an October 31 daily brief put Samsung Electronics' third-quarter HBM market share in the high 30% range, up from about 20% in the first quarter and 33% in the second — a rapid climb that helps explain why OpenAI is in Seoul this weekend. HBM4 sales more than tripled quarter over quarter and now make up about 60% of Samsung's total HBM sales, according to the same estimates, which trace to ChosunBiz reporting on October 29. The context is a market where memory is the binding constraint: SK hynix has said its 2027 HBM supply is effectively sold out, and Samsung's own shares fell 3% on October 30 even as the memory story strengthened, dragged down by the broader chip selloff. Treat the share figures as industry estimates, not company disclosures — but the direction, from challenger to near-parity in HBM within three quarters, is the story.

Source: VLSI Korea daily brief

05

Runware wants to mass-produce AI data centers like appliances

Runware plans to begin mass production of modular AI data centers in late 2027, assembling and testing its 20-foot "Sonic Pod" units in Romania before shipping them for final commissioning. Each pod packs 1,200 GPUs into 1 MW of power with closed-loop liquid cooling that needs no permanent water connection, and the company claims $1 billion in committed customer interest. A second factory is planned by early 2027, with a daily production target of 30 MW by 2028. The pitch lands in a year when Texas paused grid approvals and halted environmental permits for data-center projects: factory-built pods are an attempt to route around permitting and construction timelines entirely. The $1 billion interest figure is a company claim, not booked revenue — but the model, following Nvidia's own DSX POD reveal on October 21, shows the AI factory itself becoming a product line.

Source: Wisevoter / Cryptopolitan

06

Phinity Labs exits stealth with $5.2M to let AI agents design chips from text prompts

Phinity Labs announced October 31 that it has raised a $5.2 million seed round led by Uncork Capital and exited stealth, with a goal of automating chip architecture design using AI agents that take text prompts and iterate through a closed loop of simulation and physical feedback. The pitch is the hardware version of the agentic coding wave: instead of engineers hand-tuning architectures, agents create, simulate, and refine computational systems without manual intervention. It is early — a seed round and a stealth exit, not a taped-out chip — but it sits squarely on this month's theme, from the OpenAI–Synopsys AI-native EDA partnership to AMD and Intel's shared x86 PyTorch layer: the design process itself is becoming an AI workload.

Source: Wisevoter / Economic Times

07

A decommissioned Nvidia supercomputer lists on eBay for $59.7 million

A reseller listed a decommissioned hyperscaler cluster on eBay for $59.7 million, or best offer: 300 racks holding 9,600 Nvidia A100 40GB GPUs and 4,200 AMD EPYC processors, with 384 TB of GPU memory, about 1.4 PB of combined system memory, and 268,800 Zen 2 CPU cores — with free FedEx Ground shipping (overnight for $50 extra). The listing has since ended with no sale confirmed, and the A100, a 2020 Ampere-era part, works out to roughly $6,219 per chip at the asking price. It is a curiosity, but a telling one: even last-generation AI infrastructure commands a steep second-hand price while the industry debates how much value older accelerators retain, and every large Nvidia inventory move now draws scrutiny under export-control watch.

Source: Business Times

Cover Story · 01

Nvidia's Vera CPU holds its own against x86 in first independent benchmarks

Tom's Hardware published standalone benchmarks of Nvidia's 88-core Vera server CPU on October 30, testing a dual-socket MGX Vera reference platform against Intel Granite Rapids-AP and AMD EPYC Turin systems. Vera — built on Nvidia's custom Olympus cores for agentic AI workloads, where CPUs orchestrate tools, code execution, and data movement around GPUs — led on memory bandwidth and Python performance, edged a dual EPYC 9755 system by about 1.5% on SPEC CPU 2026 intrate, and stayed 8.6% ahead of a dual Xeon 6980P on generic GCC builds, while trailing in compression and cryptography workloads where x86 optimizations still dominate.

The result matters less as an x86 upset than as proof that Nvidia's full-platform play is credible: a CPU that feeds GPUs without becoming the bottleneck. Vera pairs its 88 cores (176 threads per socket) with up to 1.5 TB of LPDDR5X memory at roughly 1.2 TB/s of bandwidth and 1.8 TB/s NVLink-C2C links to GPUs, and Phoronix testing found near-linear scaling from one socket to two. The caveat is that these are local reference-platform tests, not cloud instances — Oracle's standalone Vera instances, generally available since October 27, remain unbenchmarked in production.

Source: Tom's Hardware

02

Intel's worst day since 1974 caps a brutal week for chip stocks

Friday's session turned Thursday's after-hours slide into a rout: Intel closed down 28.6% at $65.76, its steepest one-day drop since 1974, after its weak fourth-quarter guide landed on top of Amazon's heavy AI infrastructure spending disclosure. The selloff spread across the sector — Nvidia fell about 4.2%, AMD about 10.9%, and the Philadelphia Semiconductor Index lost 7.2% — while the Nasdaq Composite dropped 2.9%, its worst day since June. The read-through is now the market's central question: whether the trillion-dollar AI buildout earns its keep fast enough, with AMD's November 3 earnings the next checkpoint.

Source: EL7 / AP / Reuters

03

Foxconn turns Tech Day into a Vera Rubin AI-factory showcase

Foxconn opened its seventh Hon Hai Tech Day in Taipei on October 30 with 321 exhibits built around one pitch: it can manufacture the whole AI factory, not just the servers inside it. A quarter of the floor was devoted to its "Token Factory" area, including a simulated data hall showing how networking, cooling, and 800-volt DC power distribution must work together around dense AI racks, plus Nvidia's DSX POD modular reference design for factory-built Vera Rubin infrastructure and co-packaged optics work. Nvidia, Google Cloud, and Schneider Electric joined the presentations. The exhibits are demonstrations, not shipped customer systems — the test will be measured deployment times and watts at real sites — but they underline how the AI hardware contest has moved from chips to the factory floor that integrates them.

Source: BitcoinVersus / Foxconn

04

PowerLattice raises $86M to attack AI's power wall from inside the package

Vancouver, Washington-based PowerLattice has raised $86 million, according to an SEC filing on October 30, nearly tripling the $31 million it had disclosed when it emerged from stealth last November. The startup builds a power-delivery chiplet that sits inside the processor package, shortening the path current travels to the chip; it claims the approach can cut energy use by more than half, and earlier this week said it had demonstrated the industry's first monolithic integrated voltage regulator chiplet, packing controller, power transistors, and inductors onto one piece of silicon. The filing seeks up to $111 million from 18 investors, none named. The savings claim is still unverified on a customer's chip — the demo used a device built to mimic accelerator loads — but the round shows power delivery, not just compute, is now a venture-scale bottleneck, months after Analog Devices agreed to buy rival Empower Semiconductor for $1.5 billion.

Source: GeekWire

05

SK hynix rallies its Japanese suppliers around the next HBM roadmap

SK hynix said October 30 it had gathered executives from about 20 Japanese equipment and materials suppliers — including Disco, Shin-Etsu, and Tokyo Electron — in Chiba Prefecture on October 29, the first time it has held its annual Partners Day outside South Korea. The agenda: aligning production schedules for high-bandwidth memory, including commercial production of 16-layer HBM4 chips in 2026 built on TSMC logic base dies, and tool move-in for its multi-billion-dollar M15X fab in Cheongju, starting late 2026. With Samsung saying its 2027 HBM supply is already sold out and SK hynix expecting constraints through 2029, the meeting is less a ceremony than a supply-chain negotiation — the next generation of AI memory now depends on synchronizing suppliers across three countries.

Source: East Asia Brief

06

Intel says substrate shortages will cap server CPU output into 2028

A Futurum supply-chain analysis published October 30 unpacks a warning buried in Intel's earnings: a shortage of advanced organic substrates is limiting server CPU output, and the industry will stay supply-constrained into 2028. Flagship data-center CPUs have grown from 300–400 sq mm of silicon to roughly 1,600 sq mm as multi-die packages, and Intel's EMIB packaging flows — plus its capacity additions in New Mexico and Malaysia — are now part of the bottleneck, not just the fab. The analysis notes the flip side: every quarter Intel cannot ship is one AMD's EPYC Venice, Nvidia's Vera, or Arm's AGI CPU can fill, provided they can secure the same constrained TSMC capacity and substrates.

Source: Futurum Group

07

Google's space TPU phones home — with a haiku

Google confirmed October 30 that its Project Suncatcher prototype satellite — launched October 1 on a SpaceX Falcon 9 carrying four TPUs — has run its first machine-learning prompt in orbit, generating a haiku from its Gemma 4 open model during a ground-station window. Only one TPU can run at a time on the satellite's roughly 1 kW power budget, for about 20 minutes per session, and the next tests will probe radiation-induced bit flips and temperature swings in low Earth orbit. It is a research milestone, not a data center — but it is the first proof that the same accelerator silicon running in Google's data centers can switch on, infer, and transmit from space, with two more satellites planned for 2027.

Source: Interesting Engineering / Google

08

PsiQuantum pitches an all-optical switch for AI data centers

PsiQuantum, better known for photonic quantum computing, has unveiled an all-optical circuit switch aimed at AI data centers: an 8×8 prototype that, in company-reported lab tests, typically loses only 1.2 dB fiber-to-fiber — versus the 3 dB data-center optics are normally designed to tolerate — with no built-in amplification, no electrical conversion, and no moving parts, built on etched silicon circuits. The catch is scale: at 8 ports it trails Google's 136-port MEMS Palomar switch, Lumentum's 300-port offering, and Salience Labs' 32-port demo. Treat it as a technology demonstration — the continuation of this week's optics theme — not a shipping product, but one that hints at a future where light routes light without ever becoming electricity.

Source: DEV / Tom's Hardware

Cover Story · 01

Intel beats on AI servers, then warns — and chip stocks tumble

Intel's third quarter beat expectations on the back of its data-center business, but a weak fourth-quarter outlook — blamed on a prolonged PC slump and an explosion at its fab in Kiryat Gat, Israel — sent shares down more than 20% in extended trading and pulled AMD, Nvidia and Broadcom down in premarket trading on Friday. The split verdict matters because it separates the AI story from the PC story: Intel's AI server chips are selling fast, while everything tied to consumer PCs is still contracting.

Data Center and AI revenue grew 78% year over year to $7.3 billion — its strongest growth in 15 years, per the company — while client revenue fell 5% to $8.1 billion. Intel also said Nova Lake, its first client chip on the 18A-P process, is tracking ahead of schedule with roughly twice Panther Lake's design wins, with full details due at CES in January, and that its first 18A server chip, Clearwater Forest, is already shipping for revenue.

Source: Reuters · MarketBeat earnings-call transcript

02

Samsung's HBM is sold out for 2027 — so it's spending ₩84.4 trillion to build more

Samsung reported its most profitable quarter on record on October 29 and raised 2026 capital spending to ₩84.4 trillion (about $63 billion), with ₩76.8 trillion earmarked for semiconductors. The driver is high-bandwidth memory: HBM4 sales more than tripled from the second quarter to about 60% of HBM revenue, and Samsung says its 2027 HBM supply is already fully booked at agreed prices. The catch, in Samsung's own telling, is that new fabs take years to qualify — so it expects tight memory supply to persist through 2028 even at this spending pace, with HBM4E mass production planned for the first half of 2027.

Source: MLQ News · AI Affairs

03

STMicro raises its AI data-center outlook to $1.3B as silicon photonics demand grows

STMicroelectronics lifted its 2026 AI data-center revenue outlook to about $1.3 billion, up from a July expectation above $1 billion, and said the market could contribute well above $2 billion in 2027 — a management forecast, not booked sales. The engine is silicon photonics: its PIC100 platform is in high-volume production on 300mm wafers for 800-gigabit and 1.6-terabit optical modules, and ST plans to more than quadruple related manufacturing capacity by 2027. As GPUs multiply inside data centers, moving data between them optically is becoming as strategic as the compute itself.

Source: Bitcoin Versus

04

AIXTRON's AI-driven orders more than double as optical-component demand surges

German equipment maker AIXTRON reported third-quarter order intake up 122% year over year to €274.9 million, with optoelectronics accounting for about 68% of equipment orders and an order backlog of €549.4 million. InP-based optical components for AI data-center communications now make up the majority of that backlog, as the industry straddles the 800G volume platform and the gradual move to 1.6T transceivers. Equipment orders are a leading indicator: they show where data-center builders expect bandwidth demand to go next.

Source: EQS News / Sharewise

05

BrainChip puts neuromorphic AI on a Raspberry Pi 5 kit

BrainChip launched the AKD1500 Raspberry Pi 5 Development Kit on October 29, pairing a Raspberry Pi 5 with its AKD1500 M.2 board so developers can prototype always-on, low-power AI — sensing, monitoring and detection — and carry the same chip into production. The AKD1500 uses an event-based neuromorphic architecture that processes only meaningful sensor events rather than every input, and supports on-chip learning so models can pick up new classes on-device without cloud retraining. For the edge, where power budgets rule out GPUs, this is the pitch: intelligence that sips power because it mostly stays asleep.

Source: GlobeNewswire

06

Purdue brings Google TPUs to campus, lifting its AI capacity 30%

Purdue University announced on October 29 that it will add Google-designed Tensor Processing Units to its on-campus research computing system, an estimated 30% capacity increase that Purdue describes as the first deployment of its kind through a U.S. university campus computing program. Built with Google Public Sector, Cirrascale and DDN inside a renovated $16 million campus data center, the system is due for researcher use in January 2027. Purdue has not disclosed the TPU generation, chip count or price — but the model matters: custom accelerators are moving beyond hyperscalers into institutions that want local AI capacity.

Source: Bitcoin Versus

07

China's XHEART puts its 700-TOPS driving chip into mass production

Chinese AI chip company XHEART unveiled the mass-production version of its X7 Ultra autonomous-driving chip in Suzhou on October 29, claiming 700 TOPS of large-model compute — a vendor claim, measured under its own constraint that the deployed quantized model match the trained FP32 model with over 99.99% similarity. The company says the X7 series has shipped more than 200,000 units across 10 vehicle models, with orders for over 40 more, and is already planning a 3,000-TOPS X9 for 2027. It is a reminder that the autonomous-driving silicon race is now fought on large-model capability, not just driver-assist features.

Source: BigGo Finance

08

Workhorse enters modular AI data centers with a $10M TerraPod spec unit

Workhorse Group announced on October 29 a deal with UK-based Rhaelm Holdings to build a full-scale TerraPod modular data-center manufacturing spec unit at its Union City, Indiana plant for $10 million in three installments, alongside a non-binding letter of intent toward a definitive supply agreement by December 31, 2026 — including a contemplated initial order of 150 TerraPod systems. Factory-built modules that bundle power, cooling and systems control are pitched as a way to deploy AI compute faster than conventional construction; the non-binding nature of the follow-on order means the scale-up is still unproven.

Source: TipRanks