locsic.com
Thinking
Long-form notes for reading the direction of technical change.
Thinking is a professional column: fewer quick posts, more durable essays. It keeps observations, source trails, assumptions, and evolving viewpoints visible over time.
Essays
28 essays shownNo matching essays
October 2026
28 essaysWhen AI Hands Over Verifiable Proofs: Inside OpenAI's Three Scientific Claims
OpenAI announced three claims produced by an internal research model: local smoothing in all dimensions, a refutation of ETH, and Yang-Mills existence in four dimensions. The three claims shipped with three release strategies, matching three trust tiers. This analysis unpacks ETH's role in nearly thirty years of conditional lower bounds, the three boundaries of the new "subexponential but far from polynomial" algorithm, what Lean formal verification can and cannot check, and four points of contention, from the research model to the theory-engineering gap. Final validity rests with independent academic review.
Reading Anthropic's S-1: Operating Profitability Has Arrived; Compute Delivery Is the Next Test
Anthropic's Form S-1 shows GAAP operating profit of $0.974B in Q2 2026 and first-half revenue of $16.286B, roughly 14x the prior-year period. This analysis breaks down the revenue mix, earnings quality, cash flow, and compute commitments on the order of $265.6B, explains why valuing Anthropic as an ordinary subscription software company misleads, and lays out the four premises behind any valuation and the disclosures to verify next.
CMX and Distributed Memory: Which Tier Should Context Live In
Long-context inference has turned KV cache from a single-node GPU memory problem into cluster capacity planning. This article compares NVIDIA's CMX shared flash tier with distributed DRAM caching and works through a few questions: whether a restore can finish inside the deadline and beat recomputation, whether the new hits offset the cost of moving data and queueing, and whether the saved compute turns into actual savings. The 2026 memory price surge makes the last part of that math especially hard. All worked examples can be recomputed; 34 sources are cited.
DeepInfer Deep Dive: How DeepSeek Designs Model, Cache, and Scheduling into a Token Factory
From DeepSeek's open-source code and technical report: how CED, CSA2, SWA bounded replay, Wide-EP, DSpark, and async scheduling fit together; what the performance and inference-economics claims hold; and the extension interfaces, validation gates, and selection advice for production.
No New Chips at GTC: What Jensen Huang Told Berlin
GTC Berlin had no new chip: production systems stood on stage; the announcements covered factory delivery, agent runtimes, and enterprise data. NVIDIA turns the engineering problems of AI adoption into products, one by one. Customers get speed and convenience, and must reprice a new dependency on the platform.
How the Tesla Roadster Goes Airborne: Fans, Thrusters and Limits
Tesla has put tire load, electric power, underbody suction and short-duration gas thrust into one vehicle-motion control problem. This analysis checks the physics and engineering costs of fan suction, cold-gas thrusters and active aero across four evidence tiers, with a verification checklist from prototype to product.
OCP 2026: Helion and the Vera CPU Rack — Plan Agent Execution and Model Inference Capacity Separately
Plan agent execution and model inference capacity separately
OCP 2026: Open Networks Deepen — From Switch-Chip Adaptation to Endpoint Co-Design and Platform Management
From switch-chip adaptation to endpoint co-design and platform management
OCP 2026: USFP and CPO Advance in Parallel — Optical Interconnect Choice Depends on Link Bottleneck and Service Boundary
Optical interconnect choice depends on link bottleneck and service boundary
OCP 2026: Liquid-Cooling Interface Standardization — Maintenance Scope and Recovery Time Decide the Compute Hours You Lose
Maintenance scope and recovery time decide the compute hours you lose
OCP 2026: Power Standardization for the AI Factory — Power Design Must Join Disturbance Response and Service Recovery
Power design must join disturbance response and service recovery
OCP 2026: Tahan and ESUN — Ethernet Scale-up Ships First, Open Interop Still Unproven
Ethernet scale-up ships first, open interop still unproven
OCP 2026: KV-Cache Pooling Enters Validation — Returns Depend on Reuse and Transfer Cost
Returns depend on reuse and transfer cost
OCP 2026: Run:ai Goes Fully Open Source — AI Platform Builders Should Focus on Delivery and Operations
AI platform builders should focus on delivery and operations
OCP 2026: Racks Keep Getting More Open — Why Platforms Remain Hard to Replace
Chip supply and deployment capacity are now constraining AI infrastructure together. NVIDIA is widening what it opens around racks, facilities, and operations software; AMD is using Helion to organize agents' execution and inference resources; Intel is stressing capacity gains inside existing air-cooled facilities. All three compete for faster time to production — but drawings available, parts swappable, and platforms migratable remain three different things.
OCP 2026 Panorama: AI Infrastructure’s System Boundaries and Industry Opportunities
A panorama of OCP Global Summit 2026: the system boundary expands from chips to facilities and the grid, while openness diverges across racks, networking, power, liquid cooling, memory, and security. Ten chapters unpack the maturity boundaries of public signals; an appendix summarizes key items and their limits.
Inside a 10,000-GPU Token Factory: Investment, Operations and Exit
A reproducible scenario for 1,250 eight-GPU B300 servers connects rental and token revenue with costs, tax, financing and exit. Project IRR is about 7.3%; procurement prices and cache billing change the investment case.
Investing in a Token Factory: High Rental Rates Do Not Make a Good Investment
Scarcity can explain high rents today, but not returns after delivery. A 10,000-GPU reference case examines procurement timing, customer demand, token billing and financing obligations before irreversible investment.
The Token Factory Ledger: The Other Side of Soaring Demand
In 2026 the compute market raises and cuts prices at once. This article audits compute as a token factory: the token is not a stable unit of account; prices split into deflating old tiers and a firm frontier; amortization, not power, drives stock cost; China's high-end GPUs run two ledgers, with stock resale near purchase price and spot at a premium; profit settles in chips and scarce stock; pricing power anchors with model labs; the risk rests on financing and 2027 utilization.
The Network Becomes the Backplane: Arista Takes Ethernet Into the Rack
On October 7, Arista announced the SU-144 scale-up architecture: a single-hop domain of up to 144 accelerators, extendable to 1,024 across racks, using Ethernet as the backplane — alongside a liquid-cooled rack reference design. Two choices define the release. On protocol, ESUN reshapes Ethernet with a 4-byte header and link-level reliability, competing with memory-semantic UALink and proprietary NVLink. On silicon, Arista builds no switch chips of its own; the 7060XE7 line runs Broadcom Tomahawk 6. Whether scale-up revenue materializes still depends on real deployments by the open camp.
How Do You Judge Whether a Cross-Platform Ecosystem Is Truly Mature?
The testable question: when a team migrates a second model, upgrades a framework, or replaces a hardware generation, how much prior work can be reused? This piece splits reuse into four layers: interface shapes (documented alignment, capability gaps), implementation code (fully public), tests and verification methods (inheritable, verifying components not combinations), and the runtime environment (HDK not yet publicly released). The object of maintenance is combinations; exchangeable validation checklists are the route from releasing to accumulating.
Throughput Up 20%, Why Might Cost Rise?
The unit of performance comparison is not tokens per second per device, but effective work per unit of cost under the same business constraints. Kernel throughput is not service capacity: at one configuration, tokens 64 → 256 move MegaMoE from 266 to 659 TFLOPS. In this piece’s constructed example, the option 20% cheaper in resources with 20% higher raw throughput costs about 5.6% more per unit of effective output, with the break-even line near a 63.3% compliant share; one-time investment and recurring maintenance enter the payback period.
Why Can “It Runs” Still Be Wrong?
A migrated program that keeps producing output with normal latency is not necessarily correct. Two blind spots: structural error, where identical shapes and normally returning functions still pair every component wrongly (the public RoPE case); and runtime identity, where the code you think you are running and the code actually running may differ (the compile-cache signature chain). Acceptance first separates normal error from semantic error, numerical comparisons with tolerance and invariants exactly; then three layers of checks and a five-step path turn “appears to run” into “verified correct.”
Why Is FlashMLA Fast?
FlashMLA’s official 410/360 TFLOPS are sparse attention results measured under their own declared workload conditions. Behind them sit three kinds of arrangement: compute less, sparse selection reducing the attention items per query; move less, on-chip tiling and accumulation cutting intermediate round trips, where the online softmax state lives; pack tighter, the execution pipeline (64-head grouping, 1C2V, the UB-transfer supply line) deciding how fast the work completes. Each has its conditions; after the input shape changes, performance cannot be inferred from peak utilization.
Why Does MoE Get Stuck on Communication?
Measuring MoE communication by GB/s alone is not enough; four things must be tested separately: bandwidth, declining monotonically across the official five tiers from EP8 to EP128; padding, with alignment padding in this piece’s constructed example tripling processed slots; load, where a synchronous stage’s duration is set by the last device to finish; and overlap, whose gains depend on dependency structure and resource contention, measured by putting device streams, communication streams, and host-side issue/wait on one timeline.
Same API, Why Can’t It Be Dropped In Directly?
API compatibility reduces calling-layer rework; data layout decides integration cost; real matrix shapes decide performance gain; the three must be assessed separately. On Ascend, UE8M0 scaling factors are packed pairwise into int16, so mathematically equivalent computation needs a different in-memory arrangement. Under the same N and K, M=128 and M=4096 reach 549 and 861 TFLOPS. This piece also constructs a counterexample: kernel time down 30%, the whole segment 5% slower; evaluation should compare what can be handed downstream directly.
What Is DeepSeek Actually Migrating?
If you understand migration as “load the weights, run generation,” you cannot explain why this open-source batch exists. One inference batches first, then computes layer by layer: within each layer, the indexer picks context positions and sparse attention reads KV by the index; then routing, dispatch, grouped GEMM, and combine complete the expert computation; beyond API compatibility, the scaling-factor arrangement, in-kernel pipelines, and synchronization still need Ascend-specific adaptation. This piece settles the two ledgers (37B activated per token, 671B resident service-wide) and draws the reuse/redo boundary.
How Big a Training Ground Does a 10,000-GPU Cluster Need? Count Effective Samples Before Servers
Working backward from about 4.1 million trainable trajectories delivered in one week: the inference-side floor is about 74 tokens/s per card, average in-flight environments 7,100 to 13,000, folding to roughly 7 to 13 servers at the per-node peak DSec disclosed, about 16 with the 80% intake rate, about 19 if the 500-second cadence is adopted, and about 25 to plan for with margin. These numbers hold only when inference throughput, environment concurrency, and the training cadence are satisfied under the same set of operating conditions.