Powered by GIKE (General Iterative Knowledge Engine). Verified AI infrastructure signals — convergences, contradictions, and novel signals from atomic claims.https://sgnl.blog/https://sgnl.blog/2026-03-28-jevons-paradox-inference/https://sgnl.blog/2026-03-28-jevons-paradox-inference/OpenRouter data shows coding tokens grew from 11% to over 50% of all AI usage in one year. Every efficiency gain -- TurboQuant, DeepSeek Engram, cheaper models -- creates new use cases that consume more compute than was saved. The semiconductor industry is building for a demand curve that accelerates when costs drop.Sat, 28 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-26-deepseek-memory-divorce/https://sgnl.blog/2026-03-26-deepseek-memory-divorce/DeepSeek's Engram paper offloads 100 billion knowledge parameters to cheap host DRAM with only 2% throughput loss. If adopted at scale, it would double DRAM demand per AI server rack — and the DRAM shortage is already the worst in a decade.Thu, 26 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-21-hbm4-yield-game/https://sgnl.blog/2026-03-21-hbm4-yield-game/NVIDIA needs the top 20-30% of Samsung's HBM4 output to hit 10 Gbps for Vera Rubin. AMD only needs the floor bin at 6.5 Gbps. That gap isn't a weakness — it's AMD's greatest supply chain advantage.Sat, 21 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-17-ai-memory-stack/https://sgnl.blog/2026-03-17-ai-memory-stack/A single Vera Rubin NVL72 rack needs 20 TB of HBM4, 100 TB of CXL memory, and petabytes of flash — and the memory bill may exceed the GPU cost. Every layer of the AI memory hierarchy is being rebuilt in 2026. Here's the full stack, the products, and the math.Tue, 17 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-17-samsung-juggling-act/https://sgnl.blog/2026-03-17-samsung-juggling-act/Samsung is making NVIDIA's memory, fabricating NVIDIA's inference chips, developing next-gen HBM4E, and building two Texas fabs — all while its workforce moves toward a strike. One company, four critical roles, zero backup on the most important one.Tue, 17 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-16-rubin-delayed/https://sgnl.blog/2026-03-16-rubin-delayed/NVIDIA's next-gen Vera Rubin GPU — promising 5x inference over Blackwell — is reportedly delayed one quarter because the world can't make HBM4 fast enough. Google TPUs rise, AMD gets more time to close the software gap, and AI labs scramble for compute. The memory wall just hit NVIDIA's roadmap.Mon, 16 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-15-sram-inference-revolution/https://sgnl.blog/2026-03-15-sram-inference-revolution/The AI inference stack is splitting in two. NVIDIA bought Groq for SRAM. AWS rents Cerebras. But AMD already owns the deepest SRAM Compute-In-Memory IP in the industry through Xilinx — and they're the only company with GPU + FPGA/CIM + NPU + CPU under one roof. They just haven't connected the dots yet.Sun, 15 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-14-burry-vs-nvidia/https://sgnl.blog/2026-03-14-burry-vs-nvidia/Michael Burry's NVIDIA bear case has evolved from Twitter hot takes to forensic 10-K analysis. We trace his thesis through three layers: the shovel-seller narrative, the NVIDIA-OpenAI circular capital flow, and what the actual SEC filings reveal about $117B in supply commitments, permanently extending cash cycles, and hidden compensation costs. Then we stress-test it against NVIDIA's record-breaking fundamentals.Sat, 14 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-13-ai-kernels-surpass-humans/https://sgnl.blog/2026-03-13-ai-kernels-surpass-humans/DoubleAI's WarpSpeed rewrote every kernel in NVIDIA's cuGraph library and beat all of them -- 3.6x average speedup with 100% correctness. General-purpose LLMs hit only 56-59%. Here's why this matters: AI hasn't just learned to code -- it's learned to write the code that makes computers fast.Fri, 13 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-13-hardware-power-plays/https://sgnl.blog/2026-03-13-hardware-power-plays/AWS buys Cerebras for speed but not its moat. A new optical consortium draws battle lines that exclude Google and AWS. AI agents design chips overnight. And Oracle's $553B backlog is the most extreme demand-price disconnect in tech. Four stories. One thesis: the AI stack is fragmenting.Fri, 13 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-09-bottleneck-rotation/https://sgnl.blog/2026-03-09-bottleneck-rotation/For three years, GPUs were AI's only bottleneck. Now, as clusters scale past 100,000 chips and agents replace chatbots, two invisible layers are breaking: the optical links connecting GPUs and the CPUs orchestrating AI agents. NVIDIA just bet $4B on light. Intel and AMD are sold out of server CPUs for the year. The great bottleneck rotation is underway.Mon, 09 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-07-qwen35-local-agents/https://sgnl.blog/2026-03-07-qwen35-local-agents/A model that fits on a single gaming GPU now scores within 2 points of the best commercial coding AI on the hardest benchmark in the field. That sentence would have been absurd six months ago. Qwen 3.5's 35B-A3B achieves 37.8% on SWE-bench Verified Hard at 112 tokens per second on one RTX 3090. The model is ready. The hardware is ready. The software stack? That's where things get interesting.Sat, 07 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-06-token-demand-tsunami/https://sgnl.blog/2026-03-06-token-demand-tsunami/How many tokens is the world generating right now? And how many will it generate once Vera Rubin, MI455X, and the next wave of silicon come online? Nobody publishes a single answer — but by stitching together disclosed data points from OpenAI, Google, Microsoft, NVIDIA benchmarks, and shipping estimates, we can build a rough picture. The numbers suggest the industry is serving roughly 30-50 trillion tokens per day today, with capacity set to grow 10-20x by year-end. Whether demand can keep up is the trillion-dollar question.Fri, 06 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-04-anthropic-paradox/https://sgnl.blog/2026-03-04-anthropic-paradox/Anthropic's revenue doubled to $20B ARR in two months while the US government designated it a supply chain risk. An analysis of the paradox between market dominance and political crisis.Wed, 04 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-04-deepseek-v32-agents/https://sgnl.blog/2026-03-04-deepseek-v32-agents/DeepSeek V3.2 isn't just another model release — it's an architectural statement. At 685 billion parameters under an MIT license, it's the first open-weight model to unify chain-of-thought reasoning with tool-use in a single inference flow. Trained on a novel pipeline spanning 1,800+ simulated environments and 85,000+ agent instructions, V3.2 matches GPT-5 on benchmarks while its high-compute variant, Speciale, surpasses it. Here's the technical breakdown and what it means for the competitive landscape.Wed, 04 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-03-agentic-hardware-deepdive/https://sgnl.blog/2026-03-03-agentic-hardware-deepdive/The era of 'dumb' GPU clusters is ending. As AI shifts from chatbots to autonomous agents, the compute bottleneck moves from matrix math to orchestration. The CPU Pivot is reshaping data center architecture around serial logic, tool-use, and massive context capacity.Tue, 03 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-03-pentagon-deepseek-cpu-pivot/https://sgnl.blog/2026-03-03-pentagon-deepseek-cpu-pivot/15 new claims from 14 sources landed today, adding 10 graph edges. OpenAI wins a $200M Pentagon contract, DeepSeek drops V4 specs, and CPUs may outnumber GPUs in inference data centers.Tue, 03 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-01-gpu-wars-openai-mega-round/https://sgnl.blog/2026-03-01-gpu-wars-openai-mega-round/AMD secures 12GW in confirmed GPU deals from Meta and OpenAI, NVIDIA counters with 50x perf/watt on Blackwell Ultra, and OpenAI's 13GW multi-vendor strategy signals the end of single-source compute.Sun, 01 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-03-01-anthropic-lawsuit-iran-strikes/https://sgnl.blog/2026-03-01-anthropic-lawsuit-iran-strikes/Anthropic escalates to a lawsuit, US-Israeli strikes on Iran activate the oil risk premium thesis, and the tension between AI mega-valuations and bubble skeptics reaches a breaking point.Sun, 01 Mar 2026 00:00:00 GMThttps://sgnl.blog/2026-02-27-strongest-signals/https://sgnl.blog/2026-02-27-strongest-signals/AI capex is accelerating, NVIDIA's HBM moat is tightening, and the grid can't keep up. Here are the highest-conviction signals from GIKE, ranked by convergence strength.Fri, 27 Feb 2026 00:00:00 GMT