
TheValueist
17.3K posts

TheValueist
@TheValueist
Disc L/S | TMT+Energy. ISO convexity. Factor aware. Path independence matters. Results never lie. NFA. Student of mkts and cos. Creator: CRAVE Thesis of GAI.



$SMCI (Bloomberg) -- Super Micro Computer Inc. shares rose about 15% in extended trading after the server maker issued preliminary results saying its backlog hit a record on new orders in the quarter of more than $60 billion. Sales have surged for Super Micro’s servers fitted with Nvidia Corp. chips for artificial intelligence workloads. But the company has been working to get costs under control while vying with rivals to rapidly get those machines into customers’ hands and win business in a growing AI market. Super Micro said Tuesday in a statement that gross margins in the quarter ended June 30 are estimated to be in the range of 15% to 17%, which is better-than-forecast and a sign the company is making progress selling more profitable products. Fiscal fourth-quarter revenue will fall on the low end of the previous guidance of $11 billion to $12.5 billion, the company said. Analysts, on average, estimated $11.8 billion. The shares had declined 13% this year through the close, including a 28% fall on a single day last month after Super Micro announced a plan to raise $7 billion through a package of equity offerings. The company said Tuesday that the new orders “are expected to be delivered over future quarters,” which is a positive sign for future revenue and suggests that Super Micro is winning more contracts. The San Jose, California-based company is scheduled to report full quarterly results on Aug. 11. (Updates with comments from company in the fifth paragraph.)

$VICR +10% pre on raised guidance. I'm long VICR 1/21/28 c400. Power semis up across the board. (Dow Jones) -- Vicor raised its second-quarter revenue guidance due to rising product revenues and royalties from an additional licensee to its power system technology. The company boosted its second-quarter revenue outlook to $142 million from $126 million. The new license includes all of Vicor's patents covering power converter topologies, control systems, power components and distribution architectures. Shares rose 7.5% to $288 in premarket trading.



$VICR (Bloomberg) -- (Updates shares.) Vicor shares fall as much as 12% after the power equipment company reported results. Though second-quarter sales and earnings per share beat analyst estimates, they failed to impress investors after the shares more than doubled this year through Monday. SECOND QUARTER RESULTS •Net revenue $143.4 million, +1.6% y/y, estimate $138.3 million (Bloomberg Consensus) •EPS $1.04 vs. 91c y/y, estimate 63c •Income from operations $34.9 million, -23% y/y, estimate $33.2 million (2 estimates) NOTE •For Bloomberg Consensus estimates used in this story see: VICR Equity MODL •4 buys, 0 holds, 0 sells




huggingface.co/nvidia/Qwen3.6… $NVDA $MU $SNDK $LITE EXECUTIVE INVESTMENT VIEW Analysis shows that Qwen3.6-35B-A3B and NVIDIA’s nvidia/Qwen3.6-35B-A3B-NVFP4 checkpoint are best interpreted as a strategically reinforcing but not independently estimate-changing development in open-weight inference. The base Alibaba/Qwen model is a credible compact multimodal MoE release with 35B total parameters, 3B activated parameters, 262,144-token native context, long-context extensibility to approximately 1,010,000 tokens, text/image/video inputs, agentic coding positioning, and strong official benchmark results across coding-agent, tool-use, reasoning, long-context, and multimodal tasks. NVIDIA’s checkpoint is not a new NVIDIA foundation model; it is a ModelOpt NVFP4 quantization of Alibaba’s base model, released on Hugging Face on 05/28/2026, explicitly described as not owned or developed by NVIDIA, and positioned for vLLM-based deployment on NVIDIA GPU systems. The investment significance is therefore not model ownership. The significance is that NVIDIA is packaging a high-traction open-weight model into a low-precision artifact that demonstrates how Blackwell-era FP4/NVFP4 hardware, ModelOpt metadata, vLLM serving, FP8 KV cache, FlashInfer attention, Marlin MoE kernels, MTP speculative decoding, and Qwen-specific tool/reasoning parsers can be assembled into a more turnkey open-weight inference path. (Hugging Face) The correct investment conclusion is mixed and workload-dependent. Qwen3.6-35B-A3B appears meaningfully better than generic “small local model” alternatives for compact agentic coding, long-context reasoning, and multimodal enterprise experimentation, but it is not clearly a closed-frontier replacement and is not even uniformly superior to Qwen’s own dense Qwen3.6-27B sibling on every official benchmark. NVIDIA’s NVFP4 checkpoint is a meaningful deployment artifact for Blackwell-centric inference because it reduces weight memory, can improve bandwidth pressure, and gives developers an official pre-quantized path with small reported accuracy deltas versus BF16. It is not a durable standalone moat because the same base model has official FP8, Red Hat / LLM Compressor NVFP4, Unsloth GGUF and NVFP4, AWQ, MLX, and other derivatives already visible in the Hugging Face ecosystem. The durable NVIDIA advantage, if it persists, is not the file. It is the interaction among Blackwell Tensor Cores, CUDA/CUTLASS-level kernels, ModelOpt, vLLM/TensorRT-LLM/SGLang integration, NIM packaging, and enterprise support. (Hugging Face) Near-term public-equity impact is low. No credible evidence indicates that this single checkpoint changes CY2026 revenue estimates for NVIDIA, Alibaba, AMD, Broadcom, Marvell, hyperscalers, HBM suppliers, server OEMs, ODMs, networking vendors, or AI software infrastructure companies. Longer-duration relevance is higher because the event confirms 4 structural themes already embedded in AI infrastructure debates: open-weight models continue to improve quickly; inference economics are increasingly precision-, runtime-, and hardware-generation-sensitive; NVIDIA is using third-party open-weight models as a software/hardware pull-through mechanism; and non-NVIDIA alternatives remain viable because low-precision inference is becoming a broad ecosystem pattern, not an NVIDIA-only phenomenon. The most investable read-through is therefore not “Qwen NVFP4 changes the TAM,” but “Blackwell-native low-precision inference and optimized open-weight deployment are becoming a more important part of NVIDIA’s inference attach narrative, while Alibaba’s Qwen ecosystem remains a strategic asset and AMD/custom ASIC alternatives must compete on both hardware specifications and software maturity.” MODEL EXISTENCE, OWNERSHIP, AND WHAT THE NVIDIA CHECKPOINT ACTUALLY IS The official base model is Qwen/Qwen3.6-35B-A3B. The official NVIDIA checkpoint is nvidia/Qwen3.6-35B-A3B-NVFP4. The suffix A3B is material because it indicates approximately 3B activated parameters, not a conventional dense 35B model. The Qwen model card identifies the model as a causal language model with a vision encoder, trained through pre-training and post-training, with 35B total parameters, 3B activated parameters, hidden dimension 2048, token embedding 248,320 padded, 40 layers, a hybrid layout of 10 cycles of 3 Gated DeltaNet plus MoE blocks followed by 1 Gated Attention plus MoE block, 256 experts, 8 routed experts plus 1 shared expert, expert intermediate dimension 512, and MTP trained with multi-step prediction. Native context length is 262,144 tokens, with extensibility to approximately 1,010,000 tokens through long-context configuration. (Hugging Face) NVIDIA’s model card is explicit that nvidia/Qwen3.6-35B-A3B-NVFP4 is the quantized version of Alibaba’s Qwen3.6-35B-A3B model and that the model is quantized with NVIDIA Model Optimizer. NVIDIA also states that the model is ready for commercial and non-commercial use, but that the model is not owned or developed by NVIDIA and was built to a third party’s requirements. The license tag is Apache-2.0, deployment geography is global, and the stated use case is deployment in AI agent systems, chatbots, RAG systems, and other AI-powered applications. The model card lists a Hugging Face release date of 05/28/2026, architecture type “Transformers,” network architecture “Mixture-of-Experts with Hybrid Attention,” 35B total parameters, 3B activated parameters, text/image/video inputs, text output, context length up to 262K, vLLM as the supported runtime engine, Linux as the preferred OS, and NVIDIA Hopper plus NVIDIA Blackwell as listed microarchitecture compatibility. (Hugging Face) The distinction between base-model owner and checkpoint packager is central to the equity analysis. Alibaba/Qwen owns the model-quality event. NVIDIA owns the quantization and deployment-packaging event. NVIDIA’s contribution is not a new model architecture, training corpus, post-training recipe, or frontier capability claim. NVIDIA’s contribution is a pre-quantized artifact generated with nvidia-modelopt v0.44.0, tested on NVIDIA GB300, and paired with vLLM deployment commands. This makes the checkpoint important as a commercialization object in the NVIDIA ecosystem, but it limits the claim that NVIDIA has created a defensible model-layer asset. The same base model can be and already has been repackaged into FP8, GGUF, MLX, AWQ, LLM Compressor NVFP4, and other formats by official, partner, and community actors. (Hugging Face) BASE MODEL QUALITY AND CAPABILITY Qwen3.6-35B-A3B appears to be a genuinely strong compact open-weight MoE model rather than a purely cosmetic iteration. Qwen’s own model card says the release delivers substantial upgrades in agentic coding, including frontend workflows and repository-level reasoning, and introduces “Thinking Preservation,” an option to retain reasoning context from historical messages for iterative development. Official language benchmark results show Qwen3.6-35B-A3B at 73.4 on SWE-bench Verified, 67.2 on SWE-bench Multilingual, 49.5 on SWE-bench Pro, 51.5 on Terminal-Bench 2.0, 68.7 on Claw-Eval Avg, 52.6 on QwenClawBench, 29.4 on NL2Repo, 1397 on QwenWebBench, 67.2 on TAU3-Bench, 37.0 on MCPMark, 62.8 on MCP-Atlas, 85.2 on MMLU-Pro, 86.0 on GPQA, 80.4 on LiveCodeBench v6, and 92.7 on AIME26. These numbers support the view that the model is relevant for agentic coding, tool use, reasoning, and long-context workflows, but the methodology notes matter: SWE-bench uses an internal agent scaffold, QwenWebBench is an internal benchmark, Terminal-Bench uses a specific harness and 5-run average, and several agent/tool benchmarks rely on model judges or benchmark-specific scaffolding. (Hugging Face) The benchmark pattern should be treated as positive but not dispositive. Repository-level coding benchmarks are highly scaffold-dependent. Tool-use benchmarks are sensitive to parser conventions, retry logic, allowed tools, timeout settings, context allocation, and prompt templates. Frontend-generation benchmarks depend on visual judges and subjective render quality. Long-horizon agent benchmarks can be materially affected by whether the model is allowed to preserve intermediate reasoning, invoke external tools, edit files repeatedly, and recover from failed tests. The official results are sufficient to classify Qwen3.6-35B-A3B as a top-tier compact open-weight agentic model, but insufficient to classify it as a production replacement for GPT-5.5-class, Claude-class, Gemini-class, or internal enterprise-tuned coding agents without workload-specific evaluation. OpenAI describes GPT-5.5 as its smartest model as of April 2026 and positions GPT-5 around high-quality code, frontend UI generation, steerability, and long chains of tool calls, which remains the closed-frontier quality reference point for many commercial agentic workflows. (OpenAI) The 35B total / 3B active design is economically important because active compute can be far below dense 35B compute, while total model capacity can remain materially larger than a 3B dense model. This structure is attractive for inference economics because only a subset of experts fires per token. However, the cost benefit is not equivalent to running a simple 3B model. All expert weights still need to be stored or streamed. Expert routing introduces scheduling, load-balance, and kernel-efficiency overhead. MoE layers can become memory-traffic-bound rather than compute-bound if kernels are not optimized. Batch composition can affect expert utilization. The model’s hybrid Gated DeltaNet plus Gated Attention structure may reduce full-attention pressure versus a pure dense-attention architecture, but the long-context serving problem remains non-trivial because KV cache, prefill cost, multimodal tokens, and sequence scheduling dominate at high context lengths. The long-context claim is real but economically constrained. Qwen states that the model has default context length of 262,144 tokens and advises maintaining at least 128K tokens for complex tasks if OOM forces context reduction, because Qwen3.6 leverages extended context for complex tasks and thinking capability. The same model card describes ultra-long processing beyond 262,144 tokens through RoPE/YaRN-style scaling, including a vLLM example that sets max model length to 1,010,000 tokens. Qwen also cautions that static YaRN keeps the scaling factor constant regardless of input length and can hurt shorter-text performance, so the long-context configuration should be used only when needed. This is exactly the kind of caveat that matters in production: a model can support 1M context in principle while still being unattractive at 1M context under realistic latency, cost, and concurrency constraints. (Hugging Face) The multimodal claim is also real but should not be overextended. Qwen and NVIDIA list text, image, and video inputs with text output. Qwen’s official vision-language table shows strong reported performance on MMMU, MMMU-Pro, MathVista, RealWorldQA, MMBench, SimpleVQA, HallusionBench, OmniDocBench, AI2D, RefCOCO, VideoMME, VideoMMMU, MLVU, MVBench, and LVBench. The model is therefore not a text-only MoE with superficial image support. Nevertheless, multimodal enterprise deployment has a higher integration burden than text-only serving. Image preprocessing, video frame sampling, token-budget allocation, document OCR behavior, latency volatility, and vision-encoder quantization behavior can all dominate realized production quality. NVIDIA’s NVFP4 eval includes MMMU Pro, but that does not establish robust video, document, medical-imaging, industrial-vision, or long-video production quality preservation after quantization. (Hugging Face) NVIDIA NVFP4 CHECKPOINT: QUALITY PRESERVATION AND WHAT WAS QUANTIZED NVIDIA states that the model was obtained by quantizing the weights of Qwen3.6-35B-A3B to NVFP4 and that only the weights and activations of the linear operators within transformer blocks in MoE are quantized. The checkpoint therefore should not be interpreted as a naive all-tensor 4-bit conversion. The Hugging Face page lists tensor types including BF16, F8_E4M3, and U8, which is consistent with a mixed-format artifact where some components remain higher precision or are represented through auxiliary quantization metadata. The stated memory benefit is that the optimization reduces bits per parameter from 16 to 4 and reduces disk size and GPU memory requirements by approximately 3.06x. The checkpoint’s displayed model size is 19B parameters, with approximately 5,769,728 downloads last month at the time of retrieval, but Hugging Face downloads should be treated as a noisy attention signal rather than evidence of production adoption. The same page states that the model is not deployed by any Hugging Face Inference Provider. (Hugging Face) The official BF16 versus NVFP4 benchmark deltas are small. NVIDIA reports MMLU Pro of 85.6 for BF16 versus 85.0 for NVFP4, GPQA Diamond of 84.9 versus 84.8, tau2-Bench Telecom of 95.5 versus 94.7, SciCode of 40.8 versus 40.6, AIME 2025 of 89.2 versus 88.8, AA-LCR of 62.0 versus 62.0, IFBench of 62.3 versus 62.8, and MMMU Pro of 74.1 versus 74.5. The worst listed degradation is 0.8 points, 2 evals improve slightly, and the average absolute delta is approximately 0.375 points across the listed suite. This is a strong official result for benchmark-level preservation. It is not proof of production-level preservation across customer-specific coding agents, regulatory workflows, retrieval-heavy enterprise knowledge bases, internal code repositories, multimodal video workflows, or adversarial tool-use conditions. The eval table is necessary evidence, not sufficient evidence. (Hugging Face) The checkpoint’s calibration data should be treated as adequate for a general ModelOpt release but not necessarily representative of enterprise workloads. NVIDIA lists cnn_dailymail and NVIDIA’s Nemotron-Post-Training-Dataset-v2 as calibration data. NVIDIA lists training data modality, collection method, labeling method, size, and properties as undisclosed for the underlying training dataset, which is expected because Alibaba owns the base model but still relevant for model-risk governance. The evaluation suite is broader than a simple MMLU-only check because it includes reasoning, coding, telecom tool-use, multimodal understanding, scientific coding, math, long-context recall, and instruction following, but it remains benchmark-centric. Enterprise adoption will require internal red-teaming, task-specific calibration, safety review, prompt-stability testing, tool-schema validation, and regression testing under the exact runtime stack. (Hugging Face)



$VLO $MPC $PSX $DINO I absolutely agree with this. Added on top, one reasonably significant Gulf Coast hurricane and the whole sector will gap up another leg higher. I don’t believe it cuts the other way in this environment’s setup if there is no hurricane.

This is a great catch on how hard the refiners are running, the Bloomberg piece makes it clear that 95%+ utilization for months, Rocky Mountain runs over 100% and deferred maintenance are basically turning $VLO $MPC $PSX $DINO into leveraged plays on both margins and physical risk. The combination of tight inventories, storm season and gear pushed past normal limits feels like exactly the kind of backdrop where crack spreads can stay strong but a single outage or hurricane can send fuel prices and these tickers into a much sharper move than most people are positioned for. For me it argues for treating refiners as a separate sleeve in the energy stack with clear time frames and risk limits rather than just “cheap oil beta





A DM from a subscriber and it seemed like I should reply on my timeline. I’m on vacation and will share more when I’m back at my desk. Given what is happening in Iran and Ukraine, with no culmination in sight, I believe the owners of PADD 3 assets (highest concentration of total capacity) will do very well in the coming months. Crack spreads are accelerating higher. $VLO $MPC are fantastic and I like $PSX as well. eia.gov/dnav/pet/pet_p…



U.S. Refining: Structural Scarcity, Cyclical Upside, and Terminal-Value Risk. New: 7/13/26. U.S. petroleum refining is best characterized as a structurally tighter, higher-mid-cycle, lower-terminal-multiple cash-flow industry. The sector is no longer adequately modeled as the uniformly oversupplied, low-return conversion business that prevailed through much of the 2010s, but neither should recent margins be capitalized as a durable perpetuity. Capacity rationalization, limited greenfield investment, rising replacement costs, increasingly difficult permitting, high utilization, product-market fragmentation, and critical logistics constraints have raised the probability that industry troughs will be shorter and less severe than historical experience suggests. At the same time, declining gasoline demand, global capacity additions, environmental compliance costs, aging equipment, closure liabilities, and policy uncertainty reduce the duration of those cash flows and justify lower terminal multiples. High current free cash flow and declining terminal value are therefore compatible rather than contradictory conclusions.

@TheValueist Not sure what makes you think that, but fun fact: you are quoting a guy who says that 98% of his wealth is always on risky assets and that he doesn’t even use a computer to manage his positions…

Watch this video of Lloyd Blankfein, the former CEO of $GS . Ultimately, I am a risk manager and contingency planner. A key point he makes is being able to “respond very fast to what is happening because you have thought through all of the possibilities.” If you have been following since my day 1 on X, you know I am potentially the most bullish person on the GAI infrastructure trade on this app. Nothing has changed and my book reflects it (albeit now taking advantage of calendar spreads which I have shared extensively about). Given that, moreover because of that, I want to understand where ALL my risks, holes and leaks could be. I don’t want to be caught flat footed when a CEO whispers something that invalidates the GAI infrastructure trade thesis, my portfolio gaps down significantly, and I don’t know what the next move is. x.com/thevalueist/st…

Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out. Can't wait to hear what you build. Stay tuned! 🚀 Token Plan international:qwencloud.com/pricing/token-… China:platform.qianwenai.com/pricing/token-…



