The essay shelf.
Samsung's HBM3E qualification, finally. The full timeline and what it means for NVIDIA's 2026 allocation strategy.
For 18 months Samsung was the third HBM vendor on paper only. NVIDIA's quiet May qualification changes everything — including SK Hynix's pricing power and Micron's growth ceiling.
↳ Energy · GridThe hyperscalers built nuclear deals nobody can finance. Here's what happens when reality hits.
SMR contracts announced in 2024 and 2025 looked clean. The capex curves don't pencil out, and three of seven announced deals are quietly behind. Source-by-source breakdown.
↳ Policy · GeoCommerce's HBM3E export guidance is narrower than reported. The HBM4 ambiguity is the real story.
Reading the actual rule text — not the headlines — reveals three exemptions and one critical omission. We walk through every line that matters.
↳ AI · MemoryWhy every AI training run is now a packaging negotiation.
The bottleneck no longer measured in GPUs but in CoWoS slots. How TSMC quietly became the gatekeeper of the entire frontier-AI roadmap — and what the 2027 packaging map looks like.
↳ AI · MemoryCerebras WSE-4 is generally available. We ran the benchmarks. The numbers are real.
Wafer-scale inference has spent four years as a press release. With WSE-4 in our hands, the latency-vs-cost economics finally pencil out for a specific, narrow class of workloads.
↳ Cloud · InfraThe neoclouds are losing money slower than expected. CoreWeave's S-1, two years on.
Margin compression was the consensus call. Public filings tell a different story — driven by HBM allocation tactics and one very specific contract clause.
↳ AI · MemoryWhy Samsung's HBM gap is closing.
Samsung's HBM3E yield rates climbed faster than expected in early 2026. We look at what changed in Samsung's production process, the NVIDIA qualification milestone, and whether the third-vendor story is finally real.
↳ AI · MemoryMicron, year of the ramp. The "American HBM" trade is now structural, not speculative.
Micron's HBM3E sold out through 2027 and the HBM4 sample timeline is closing the gap with SK Hynix. What this means for supply diversification and 2026 pricing dynamics.
↳ Silicon · PackagingInside TSMC's CoWoS expansion plan — and why HBM supply still won't catch up.
The newly announced $3.8B in incremental capex moves the needle, but not until 2027. Until then, packaging is the hard ceiling on HBM-bonded GPU output.
↳ AI · MemoryPaged attention, two years later.
vLLM's paged attention paper changed inference economics overnight. Two years of production data later, here's what actually happened to the KV-cache problem — and what's still unsolved.
↳ AI · MemoryThe FP8 → INT4 quantization roadmap.
Inference vendors are racing from FP8 to INT4 as the next lever on compute efficiency. We map the roadmap across frameworks, the accuracy tradeoffs that actually matter in production, and which memory vendors benefit most.
↳ AI · MemoryHBM, explained without metaphors. Stacks, TSVs, and why it's so hard to make.
High-Bandwidth Memory isn't "fast DRAM." It's a wholly different manufacturing problem. We go through the stack, the through-silicon vias, and the packaging step that ties it all to the GPU.
↳ AI · MemoryThe HBM4 spec war: how a memory standard became an NVIDIA-customer fight.
JEDEC ratified the standard. NVIDIA's customer requirements then re-opened it. We walk through the spec deltas, the supply implications, and which vendor benefits most.