beerbaron

Exactly, the cost of software execution went down. However, there is now additional need for upfront weeding out the ideas as those failed trys will go nowhere and put strain on other areas of organization. Example, sales do not have infinite bandwidth to sell new stuff. If new features are not sold and implemented without customer knowing... Why are you doing it. Lots of waste because people want to skip the very essential task of thinking and debating the why and the how. My recommendation to clients i meet that want to start with ai is usually. Identify an operation that is Major pain point within your business and implement POC WITH MEASURES of success before implementation.

I think these examples are mistaking the trees from the forest. This is just a function on how LLM operates. Does not show lack of intelligence but rather how it see things (token).

AMAZING. This is a model on par with top end frontiers from december 2025. 1/15th of the price of opus. Imagine if China pulls out a better coding frontier model just before the 3 big IPOs, that would cratter valuations. How many over here know if there company allows dev. with chinese models? I

Does that Mythos thing remind you guys about when Japan was supposedly banning PS2 exports because it could be used in missile? Seems like a bit of marketing to me.

What is your issue with chinese models?

Well sentiment has little to do with investment in us assets. Based on economic accounting principles and the data shown in tbe graph, the link between increasing foreign investment and the trade deficit is direct and functional: they are two sides of the same coin. Fix the trade deficit and foreign asset will follow. Fixing trade deficit will not be done by tarifs alone. Fiscal policy and consumption reduction will do the heavy lifting.

The below is another threat to consider too. Not AI directly related... but will be a trend. Greenland was a breaking point, the decoupling will take many years but the will is now there. https://youtu.be/oZL-nRB23RI?si=cluwr16q6Hyv_lui

I think the AI folks are trying to sell a dream here. There are many layers of what we can call an app. I asked Gemini to give me a framework to help start the tough process. I think it can be refined by humans tough. 1. Structure of Data The Question: Does the company rely on data that anyone can scrape, or do they own a proprietary, "closed-loop" dataset that AI cannot access elsewhere? High Risk: The software processes public information or standard business documents (resumes, basic legal contracts, generic code). Low Risk: They have 10 years of private, industry-specific telemetry or "human-in-the-loop" feedback that makes their AI significantly smarter than a general model. 2. Complexity of Workflow The Question: Is the software a "system of record" (where data lives) or a "system of effort" (where work happens)? High Risk: The tool's main value is saving time on a single task (e.g., "AI copywriter" or "background remover"). These are features, not companies. Low Risk: The software is deeply integrated into a messy, multi-step business process involving multiple departments, permissions, and legacy hardware. 3. Output vs. Interface The Question: Does the user care about the software's buttons, or just the final result? High Risk: The UI is a complex dashboard that takes weeks to learn. An AI-native competitor could replace 50 buttons with one "chat" box or an agent that does the work in the background. Low Risk: The interface is the value (e.g., collaborative design tools like Figma or real-time communication tools) where human-to-human interaction is the point. 4. Revenue Model (Pricing) The Question: Do they charge "per seat"? High Risk: If a company charges per user, and AI allows 1 person to do the work of 10, the company’s revenue will drop by 90% even if the customer is happy. Low Risk: They use Value-Based Pricing (charging based on outcomes or usage) rather than the number of human logins. 5. Ecosystem & Integration The Question: How hard is it to "unplug" them? High Risk: It’s a browser extension or a standalone app with no deep integrations. Low Risk: They are the "Single Source of Truth." If turning off the software breaks the company’s payroll, taxes, or supply chain, they are safe (for now). Comparison: High Risk vs. Low Risk

Here is what I predict will happen next: TRUMP : "If you resist an arrest you deserve to die" Of course... totally normal.

Well, as a Canadian I just hope Carney had an intent behind it's speech and that escalation was somewhat part of the plan. 4D chest game... we shall see in 3 months. BeerBaron

I dont understand why its being discussed here. Fraud is not a political subject and has nothing to do with any sides of isles. It has to be caught and fixed. Nothing to debate. I find it irrelevant if the fraud was from immigrants or a sub contractors for the US Air Force. Lots of things should not be politized and this is one of them.

Yeah I was thinking about the lasers, optical transducers, etc... I don't think those have reached the performance ceiling equivalent to the silicon limit. I'm a bit wary of investing in anything related to AI hype right now but if your are going to scale super-horizontally you need ridiculous latency and bandwidth to avoid bottlenecks. We should start another thread discussing opportunities for picks ans shovels for AI where valuations have not hit stratosphere yet. BeerBaron

NVidia does does disclose MTBF but it's mentioned that they have a median life around 6 years. The problem with GPU is that it's all parallel work. So if you lose a piece, the whole round of compute might be lost. In a way it's kind of the same thing as wireless networks (packet collision when two client emit at the same time), below a certain level and it is quite manageable but above another limit and the whole thing becomes useless. You can obviously mitigate that by mixing compute between GPU or having some redundancy which is what everybody is doing but then, you add a load on the telecom side of things. In fact as I come to think of it, I'd rather invest in some next gen telecom for GPU cluster than the GPU themselves. Demand should scale as much but with a much better reliability and still has room to improve for quite a while. BeerBaron

Well, if you listen to Jensen's interview on BG2 he talks about how B100 VS H100 and puits some 2X number and much higher when optimized to the new hardware. According to him, it make absolutely no sense to buy H100 if you can buy B100. The operating cost of the H100 VS the B100 is much higher per TFLOPS. So one could argue that if there were not supply shortage those old GPU would have a negative value. Similar to when miner ASICs were profitable until an ew generation came up... they became fancy space heaters over night and made more NPV to thrown then in the trash than keeping operating them. I would say if NVIDIA scales up production it basically destroys the amortization period. Furthermore, amortization is currently artificially long because of a supply issue and that if your are not a hypercaler with long terms commitments you are shit out of luck. I asked ChatGPT to make me a comparative table comparing the same TFLOPS (reference TFLOPS is 100xH100 cluster) between architectures. TAKE WITH A GRAIN OF SALT AS I'M NOT AN NVidia ANALYST. But it gives and idea. Cost Component A100 Cluster (600 GPUs) H100 Cluster (100 GPUs) B100 Cluster (56 GPUs) GPU CapEx $9,600,000 $3,000,000 $1,820,000 Server / Node Infrastructure (CapEx) $6,000,000 $1,000,000 $1,400,000 Total CapEx $15,600,000 $4,000,000 $3,220,000 --- --- --- --- 5-year Power Cost $1,365,000 $398,500 $223,200 5-year Cooling / PUE Overhead (included in power) (included) (included) 5-year Maintenance / Spares $1,330,000 $450,000 $322,000 5-year Staff / Operations Cost $5,500,000 – 6,000,000 $1,000,000 – 1,300,000 $800,000 – 1,100,000 Total OpEx (5 years) $8,200,000 – 8,700,000 $1,850,000 – 2,150,000 $1,345,000 – 1,645,000 --- --- --- --- 5-Year Total TCO $23.8M – $24.3M $5.85M – $6.15M $4.56M – $4.87M

Impossible to overclock GPU in a large cluster, they would all have to be equally overclocked in order to see the gain. There is kinda of a sync period where between each propagation where data is more or less shared across the whole GPU cluster. Furthermore, GPUs are notoriously unreliable, losing a few GPU can significantly delay the training process, imagine if an overclock triples the early death of those GPU.