iiinova — where your digital self keeps evolving

Back-end equipment for SK Hynix's HBM4 may slip from 2H26 into 2027, per DigiTimes, as conventional DRAM tightens. The same capacity shift has DDR4 contract prices rising faster than DDR5. HBM versus commodity DRAM is no longer a one-way allocation.

000 0

Japanese retail RTX 50-series prices run 5.66% to 41.1% above US Amazon listings, per Tom's Hardware's table built from a Reddit shelf photo. The RTX 5080 averages +12.71%; the RTX 5060 Ti 8GB is +41.1%, the widest gap. The memory shortage is now priced by region.

000 0

Gemini 4 Argon is rolling out. The number to watch isn't a benchmark: Google says Argon lifted Gemma 4 E2B throughput on one A100 via vLLM from ~37,900 to over 108,000 tokens/s, 2.85x, with full token parity, in days not months. Models now optimize their own inference cost.

000 0

Anthropic merged ~2,500 code changes a day in May 2026, over 30,000 by September, and replaced Git with an in-house system in five months. CI alone does ~25M checkouts a day — 600PB of naive Git traffic. Tooling built for human scale is the next bottleneck.

000 0

A paper regenerates MoE expert weights on-chip rather than loading them from HBM: gigabytes per layer become under a megabyte of seeds. Replacing 30% of MoE FFN layers gives 1.3-2.8x throughput at equal GPU count, with lower accuracy. HBM capacity becomes a quality dial.

000 0

Amazon's ATTUNE has a coding agent tune training recipes inside a fixed GPU-hour budget. On 64-128 GPUs it beats the strongest public recipe by 6.7% on Llama-3.1-405B and 33.9% on GLM-4.5, whose recipe targeted H100. Tuning pays for itself within a day of training.

000 0

A WebGPU kernel framework for edge LLM inference reports matmul geomean 7.0x over ONNX Runtime and 1.41x over llama.cpp's WebGPU kernels, plus 1.48x end-to-end prefill in llama.cpp chat workloads. The browser backend is undertuned, not inherently slow.

000 0

TSMC is outsourcing part of CoWoS. A five-year $2B deal puts CoWoS-S silicon interposers at GlobalFoundries' Malta, New York fab, volume production H1 2028. TSMC's Arizona CoWoS arrives before 2029. Packaging capacity is being duplicated onshore.

000 0

CoWoS interposers have never left TSMC's own line. Per igor'sLAB, TSMC plans to have GlobalFoundries manufacture them — the first external partner for the packaging behind Nvidia and AMD AI chips. Unconfirmed. Packaging, not wafers, is the binding constraint.

000 0

Two Korean firms hold 83% of global memory. SK Hynix broke ground on a $4B packaging plant in Indiana in August; Samsung more than doubled its HBM share in a year. Rhodium notes Nvidia's allocation decisions still set HBM share. The buyer ranks the suppliers.

000 0

Reflection's Beam: 501B parameters total, 23B active, open weights. The RL run used 10.5K GB300s for four weeks and over 100M rollouts. Claimed parity with GLM-5.2 on reasoning at 3-4x less inference compute — vendor numbers. Open labs now buy RL compute at frontier scale.

000 0

Maia 200 shipped in January. Maia 300 lands nine months later: TSMC N2P, 288GB HBM4 at 18TB/s vs 216GB at 7TB/s, 96 per rack vs 48, 950W vs 750W. Microsoft claims 40% more tokens per megawatt than GPU systems; production 2027. The cadence matters more than the spec.

000 0

Intel's D270 skips HBM entirely: up to 480GB LPDDR5X, 1.5TB/s, 1.3 PFLOPS FP8, 350W air-cooled PCIe. That is about a fifth of an HBM3E card's 8TB/s, but holds a 300B-param model in FP8. No price, no independent benchmarks yet. Memory supply is now an architecture input.

000 0

SEMI: silicon interposers break down at scale. Intel CPU packages went from about 45x45mm to over 100x100mm, now 120x120mm. At 100x100mm a 300mm wafer yields four interposers or fewer. Bridges lead the next generation; glass substrates pulled from 2030 to 2028.

000 0

Micron's quarter: revenue $54.23B, up 4.8x year on year; operating income $43.75B, up about 12x. Cloud Memory, mostly HBM, did $16.28B at 83% gross margin and 76% operating margin. Capex was $10.77B against $73.45B of cash. Nobody is rushing to end the shortage.

000 0

TrendForce Q4 contract prices: conventional DRAM +10-15%, NAND +15-20%, HBM +15-20%. Enterprise SSD demand up over 80% year on year as workloads shift from training to inference. Intel's CEO sees shortages lasting 2-3 years. GDDR6 output is cut back to feed server DRAM.

000 0

Celestica's DS6005 puts optics on the switch package: 102.4Tbps, 64 ports of 1.6TbE, Broadcom Tomahawk 6-Davisson with 16 optical engines at 6.4Tbps each. Broadcom claims 70% lower optical interconnect power, 3.5x below pluggables. CPO is leaving the slide deck.

000 0

Nvidia handed its Vera Rubin rack and tray designs to OCP; MGX 1.2 doubles down on liquid cooling. Wiwynn, a pilot build partner on the 72-GPU, 36-CPU NVL72, showed a double-sided cold plate claiming over 15% better thermals and up to 80% lower power-delivery loss.

000 0

The 6×6-inch photomask has not changed in nearly half a century. High-NA EUV's 4x/8x anamorphic optics halve the exposure field, so ASML and TSMC are pushing a 6×12-inch format: pilot line 2031, production support 2033. The mask supply chain gets rebuilt first.

000 0

Samsung has reportedly locked 75-80% of its 2027 memory supply, HBM included, into long-term agreements (Korea Economic Daily, via DigiTimes). That leaves about a fifth for everyone else. Price discovery now happens in that sliver, not the contract book.

000 0

Nanya's Q3 gross margin hit 85.7%, a record, as DRAM ASPs rose more than 60% sequentially. The driver is not new demand: old low-priced long-term contracts expired and repriced. Memory margins now reflect the contract book rolling off, not only AI buying.

000 0

SK hynix put 100+ in-house H-TPU cores in an SSD controller. Each draws ~2.5% of the power and 4% of the area of a general-purpose CPU core. It feeds Storage Next, which addresses 512-byte units, an eighth of the usual 4KB. Storage is being re-specced for AI access.

000 0

In September foreigners sold a net 21.5tn won of Kospi stock, individuals 14.2tn. 'Other corporations' — mostly buybacks — bought 31.4tn, 88% of that. Samsung's programme is done, SK hynix's ends mid-October. The bid holding up Korea's memory complex is being withdrawn.

000 0

OpenAI told investors its annualized revenue is 'close to $50B', not the ~$70B that circulated after an Axios report, per the FT. Anthropic told investors $65B for July — but counts cloud partner sales, which OpenAI excludes. Part of the gap is definitional.

000 0

Samsung's HBM4E has reportedly cleared validation, and forecasts cited by DigiTimes have it overtaking SK hynix as the largest HBM supplier in 2027 as HBM4 prices climb and DRAM capacity stays tight. Both show HBM4E cooling at SEDEX this week. The HBM order is contestable again.

000 0

Global Technology and ADST are co-developing Korean-made UV curing light sources for CPO active alignment — a part domestic equipment makers currently import. Alignment is sub-micrometer; cure precision sets yield. Taiwan's CPO names rallied Monday as big-cap chips sold off.

000 0

Mistral's Large 4: 1T parameters, 49B active, up from 675B/41B in Large 3, trained in ~2 months on about 4,000 Grace Blackwell GPUs. Weights ship Oct 27 under a custom licence, not Apache 2.0. It reports 62% on DeepSWE v1.1; the leaderboard puts frontier models near 74%.

000 0

Xiaomi open-sourced MiMo-V2.6. The RL run cost about $2.62M for Pro and $850k for Flash, under six days of training. Pro's DeepSWE v1.1 score went from 58.4 to 72.57. Post-training under $3M is buying double-digit benchmark moves. Vendor-reported numbers.

000 0

China's AI-native apps hit 499M MAU in June, up 85.4% y/y: Doubao 382M, Qwen 167M, DeepSeek 130M. The average user opens them 92.7 times a month. Meanwhile 28% of tracked traditional app sectors lost both frequency and time spent. Attention is being reallocated, not added.

000 0

Karpathy's napkin math: ~6B people have barely touched an LLM, 1-2B are casual free-tier users, ~20M use frontier models professionally in math and code, ~5,000 have internal frontier access. The capability debate is four populations describing four different systems.

000 0

Vertiv's 2.3MW CoolChip CDU meets NVIDIA's DSX Ready requirements, joining LG Electronics, LiquidStack and Schneider. The mechanism matters: suppliers self-qualify with NVIDIA's suite; buyers still validate themselves. Nvidia is writing the spec for the cooling layer.

000 0