Rocky_BTC

Long term investor #BTC #TAO #SOL #SUI #XRP #OKB| MeMe Professional Data Player | Crypto since 2017 | Not financial advice, DYOR🙏Twitter：@Rocky_Bitcoin

5Following

209followers

Feed

#OpenAI hasn't gone public yet, but its computing power subsidiary is about to IPO, which is quite interesting.🧐 Previously, Jensen Huang mentioned that inference demand will grow by a billion times in the future. Next Thursday, May 14, the dark horse AI chip of the inference era, Cerebras ($CBRS), will go public with a price range of $115-$125, aiming to raise up to $3.5 billion, with a valuation of $26.6 billion. It will also be the first Pre-IPO project of @MSX_CN, which is very much worth looking forward to! Today, let's break down this Cerebras company, analyze its valuation, and share some of my personal judgments and opinions. To understand this OpenAI-related inference chip dark horse, you need to know Sam Altman's capital layout. We all know how powerful NVIDIA is in the AI chip field: large model companies burn money, cloud providers buy cards, startups line up for GPUs, and most profits flow to NVIDIA, the shovel seller. This is the current industry status. But in this monopoly situation, all major large model companies hope to have a loyal Plan B solution. For example, Google Gemini partnered with Broadcom to use TPU solutions, and OpenAI has always wanted to support its own loyal forces. So on May 6, OpenAI brought together NVIDIA, AMD, Intel, Broadcom, Microsoft—chip companies that should be competitors—to create the MRC network protocol. On the surface, it looks like technical cooperation, but actually, OpenAI wants to redistribute the pie. Looking deeper, I believe #OpenAI wants to break NVIDIA's full-stack monopoly. Previously, training, inference, networking, and cloud were all monopolized by NVIDIA. Now? OpenAI is starting refined operations: training for training, inference for inference, using different chips for different scenarios, and different suppliers for different stages. #Cerebras was pushed to the table at this time, responsible for the inference part. This coincides with the recent hype around inference CPU concepts, such as #AMD, #INTC, etc., hitting the hot spot. 🔥What makes Cerebras so powerful? Cerebras's core trump card is the WSE-3 chip, which directly makes an entire 12-inch wafer into a giant chip, with an area of 46,225 square millimeters, about one-third the size of an A4 sheet. Let's compare data with NVIDIA's H100: • Its area is 57 times that of H100 • Core count is 52 times • On-chip memory is 880 times • Memory bandwidth is 7000 times These numbers 📊 seem exaggerated, but the key is not size, but speed. In inference scenarios, especially long text output, real-time interaction, code generation, AI Agents—tasks requiring low latency—Cerebras's CS-3 system inference speed is 21 times faster than NVIDIA's DGX B200, with cost and power consumption reduced to one-third. This efficiency and power consumption mean that with the WSE-3 chip, OpenAI can serve more customers per unit time, which is like pure money saved. 📊Financial data is also impressive From the market trend perspective, the AI industry is shifting from training-dominated to inference-dominated, an indisputable fact. The global AI inference market size will reach $106.2 billion by 2025 and is expected to grow to $255 billion by 2030. Cerebras's technological advantage is right on this trend. Additionally, this IPO round values the company at $26.6 billion, with an issue price of $115-$125 per share. I think this is relatively cheap, although it doubled from the previous round's valuation of about $12 billion in the F round. In just two years, it doubled, but the financials are impressive. Cerebras's revenue in 2025 is projected at $510 million, a 76% increase from $290 million in 2024. Even more impressive is the net profit of $87.9 million, turning from a $485 million loss in 2024 to profitability. At a $26.6 billion valuation, the PS ratio is 52x. Compared to the hot semiconductor interconnect chip company Astera Labs (#ALAB) that went public in 2024 with a PS of 81x on the first day, and given the current hype in the inference track, I personally believe Cerebras can easily reach 80-100x PS, corresponding to a closing price of $192-$239, expecting over 50% upside! (But also need to consider the Nasdaq index on that day for a comprehensive judgment.) Not only the positives, Cerebras also has obvious issues: customer concentration is too high. MBZUAI from UAE contributes 62% of revenue, G42 contributes 24%, with the top two customers accounting for 86%. This means Cerebras must listen to big customers, limiting autonomy. Fortunately, with OpenAI's involvement, this revenue structure will improve, and OpenAI will become the largest customer. 🎯Deep binding between OpenAI and Cerebras Latest data shows OpenAI and Cerebras signed a multi-year cooperation agreement worth over $20 billion. Cerebras will provide OpenAI with 750 megawatts of computing power, deployed through 2028. But this is not just a procurement contract. OpenAI's founder Altman, president Brockman, former chief scientist Ilya, and board member Adam Angolo—all core executives—have personally invested in Cerebras. OpenAI also uses loans, warrants, and other financial instruments to establish long-term interest alignment with Cerebras. Simply put, Cerebras is now OpenAI's chip division. Besides, in March, Cerebras partnered with AWS to launch the CS-3 system on Amazon cloud services, becoming the first non-GPU AI accelerator to enter the mainstream cloud provider supply chain. Also, clients include GlaxoSmithKline, the U.S. Department of Energy, and multiple national labs, validating its technical strength from multiple dimensions. 💡OpenAI's capital strategy OpenAI's real intentions are clear: • Continue using NVIDIA high-end GPUs for training • Introduce Cerebras low-latency solutions for inference • Purchase some GPUs from AMD • Open network protocols • Bet on multiple cloud services: AWS, Azure, Google Cloud • Possibly develop self-designed chips in the future This is a computing power combination strategy, matching different workloads with different systems, no longer relying solely on NVIDIA's full-stack solution. OpenAI is transforming from a model company to a computing architecture company. Previously, it passively accepted chip vendors' technical routes; now it actively designs computing power combinations that fit its needs. OpenAI wants to downgrade chip suppliers from "platform providers" to "module suppliers." Supporting Cerebras is the most important part of this strategy. Looking ahead, Cerebras's IPO day stock price has a very high chance of exploding! ⚡Impact on NVIDIA In the short term, Cerebras's IPO won't significantly impact NVIDIA, like a minor pimple on the body. NVIDIA currently holds 80-90% of the AI chip market share, with CUDA ecosystem, GPU supply chain, NVLink network—these moats are hard to shake in the short term. But in the long term, the threat exists. Previously, AI companies had no choice but to use NVIDIA GPUs. Now, at least in inference scenarios, customers have viable alternatives. This choice weakens NVIDIA's pricing power. When OpenAI can say "I use Cerebras for inference, NVIDIA for training," NVIDIA loses its "all-in-one" bargaining power. The AI inference market is growing rapidly. Forecasts show a 28.9% CAGR globally from 2026 to 2032. Inference scenarios are more suitable for specialized chips. When the inference market surpasses training, NVIDIA's relative weakness in inference will become a bigger problem. NVIDIA is shifting from "sole supplier" to "one of the core suppliers." This change is not because NVIDIA is weaker, but because the market is bigger, customers are stronger, and demands are more complex. 🧐My judgment What’s truly worth watching about Cerebras's IPO is not just another AI chip company going public, but that OpenAI is starting to price inference as a separate business. When the inference market is validated as independently priced, the AI computing power market truly begins to stratify. The demand differences between training and inference become clear, and specialized chips prove their advantages in segmented scenarios. NVIDIA's narrative of "one chip rules all" no longer fully holds. The market will move from "general GPU monopoly" to "scenario-based chip combinations." And it's not just OpenAI doing this; Anthropic is also allying with Amazon and Google. Leading AI companies are diversifying procurement to reduce dependence on NVIDIA. A single supplier's "complete solution" is no longer the optimal choice. Finally, worth mentioning is that this #MSX first Pre-IPO project is about to go public—it's Cerebras this Thursday. Stay tuned!🧐 DYOR🙏

Over the past decade, from an investment perspective, I've gradually come to understand one thing: companies that build Apps eventually get crushed by competition; the ones quietly making money are those collecting "toll fees" as landlords. 🧐 Look, back in the early internet days, everyone wanted to be the next Taobao or WeChat, but what happened? Most of them failed. Instead, essential services like Alibaba Cloud and Tencent Cloud, which sell servers, made money effortlessly. The Web3 era is the same—99% of DApps and NFT projects went to zero, but Ethereum’s Gas fees kept coming. I believe the AI era will follow a similar logic! Based on the chart below 👇, I want to talk about where the opportunities lie in the next 10 years and some core US stocks and crypto investments we currently hold! The fundamental first principle is simple: when investing in Web3 and AI, minimize investment in the application layer and focus on underlying infrastructure and essential toll-collecting layers. The application layer is exhausting—some projects spend half a year creating a hit, only to be copied within three months, and they have to spend daily on traffic acquisition. But the toll-collecting layer is different. As long as people need internet access, computing power, and transactions, it collects money passively. Applications can change, but the underlying "toll fees" are unavoidable. Data from the past decade shows: Nvidia rose 221x, Ethereum 236x, Bitcoin 178x. • Nvidia collects the "AI computing power tax"—if you want to do AI, you have to buy my cards, no choice. • Ethereum collects "Gas fees"—if you want to transfer on-chain or play with NFTs, you must pay fees. • Bitcoin collects the "consensus tax"—everyone recognizes it as digital gold, making it an asset itself. See the pattern? The real big money is made by those who control the "tax rights." 📊 For the next 10 years, I am optimistic about these three tax-collecting sectors and related projects and companies. 1️⃣ The "interest spread tax" of digital dollars (stablecoins and RWA) To play DeFi or invest in RWA on-chain, you first convert money into USDT or USDC. Issuers use our dollars to buy government bonds and earn interest, giving us a digital token. The larger the scale, the better the interest. This is a guaranteed profitable business. Especially with the coming AI Agent era, USD stablecoins are the natural payment medium, with transaction volumes expected to reach trillions. This is a solid "interest spread tax"—whoever controls this gateway is the Nvidia of finance. Key picks here: • Circle (US stock ticker #CRCL), as the compliant stablecoin USDC, is currently unique. When institutions want to put trillions of assets on-chain, compliant stablecoins are the only gateway. With no rate cuts expected this year and high US Treasury yields, stablecoin issuers will have more stable earnings. Circle is also aggressively pushing to capture AI Agent payment gateways, holding a 68% market share, making it the preferred stablecoin for AI Agent payments. • Ondo Finance (crypto ticker #ONDO), a leader in the RWA sector. It brings tokenized US stocks and bonds on-chain. Still in the "early paradigm" stage. From a tax logic perspective, as on-chain RWA assets grow, ONDO, as a liquidity manager, will take a cut from every on-chain US bond and stock yield. It will become the "tax collection gateway" connecting traditional and crypto finance. The only downside is weak token empowerment, lacking buyback and deflation mechanisms! 2️⃣ The "physical essential tax" in the AI era (power and energy supply chain) AI ultimately depends on electricity. No matter how powerful the chips are, they won’t run without power. Data centers are power-hungry monsters. I focus on nuclear power revival (SMR small modular reactors), transformers, and liquid cooling equipment. Without power, even the strongest AI is useless. This is a hard tax—inescapable. Key picks here: • Vertiv Holdings (US stock ticker #VRT), a company we recommended early, trading around $100. It’s a global leader in data center liquid cooling and power management. AI chips generate enormous heat; traditional fans can’t keep up, so liquid cooling is essential. Vertiv dominates the high-end liquid cooling market. From a tax perspective, it’s an indispensable part of data center construction. Without its cooling systems, expensive NVIDIA chips would burn out. • OKLO (US stock ticker #OKLO), a leader in SMR small modular nuclear reactors. Most SMRs worldwide are still in design and approval stages, but OKLO has received site approval from the US Department of Energy and is getting closer to commercial operation. It’s an early paradigm company with deep technical barriers. OKLO uses fast reactor technology that can use nuclear waste as fuel, offering high compliance and environmental barriers. It has investment and endorsement from OpenAI founder Sam Altman, showing huge potential. 3️⃣ AI Agent identity/distribution/payment layer, the "Apple tax" of the AI era In the future, AI—not humans—will do our work. AI will book tickets and buy things; it needs identity and direct payment capability. Whoever sets the payment standards and identity protocols between AIs will control the AI era’s throat. I see this as the next breeding ground for 10x tokens. Key picks here: • World Network (crypto ticker #WLD), in an era full of AI fakes, proving "I am human" or "this is my data" becomes extremely scarce. #WLD is a digital identity protocol founded by OpenAI’s Sam Altman, using iris scans for real-person verification and uniqueness. Imagine when 90% of future internet content is AI-generated, an account proving you’re human is the hardest ticket. With OpenAI’s IPO this year, the opportunity is obvious. Tax logic: if all social media and banking apps require World ID login, it becomes the global digital world’s "entry toll." • Bittensor (crypto ticker #TAO), like the "Taobao of AI." Whether you generate images, translate code, or scrape data, you can open a "subnet stall" there. All subnets must stake TAO to operate, and all computing contributions and intelligence evaluations are settled in TAO. Anyone wanting to call decentralized AI capabilities through this network must use TAO as the settlement currency. Bittensor today is like early Ethereum. Ethereum collects "financial tax" via smart contracts; Bittensor collects "intelligent transaction tax" via the Yuma consensus mechanism. Although early subnet supporter Covenant AI’s founder Sam Dare recently left, causing some turbulence, it doesn’t affect the project itself. We continue to watch its subnet ecosystem breadth, which is TAO’s core. Besides the above, in crypto, we also hold infrastructure like oracles and public chains—#LINK, #PYTH, #SOL, #SUI, etc. Whether for prediction markets or AI Agents, real-world data like asset prices, weather, and event info are needed, making them essential tax-collecting projects! Public chains are even more obvious—Web3 apps must be built on them, making them long-term toll collectors! 🎯 Recently, we’re also watching two hidden sectors: • AI model evaluation and auditing—regulation will surely govern AI. Who proves AI is safe and unbiased? This "trust tax" will become increasingly valuable. • Edge device gateways—not all computing power will be cloud-based. Phones, glasses, and headphones will run AI locally. This "terminal gateway tax" could be a hardware maker’s comeback opportunity. For example, Anker’s integrated compute-storage chip Thus provides AI local computing units for headphones. 🧐 How to pick 10x targets? The formula in the chart is quite practical: 10x opportunity = Real tax layer × Early paradigm × Hard to bypass × Undervalued Find early-stage—don’t chase what everyone knows; look for emerging sectors the public hasn’t understood yet (like AI Agents, nuclear power, AI chips focused on edge devices). We’re already starting to take profits from memory chip companies like #MU. Check the moat—ask yourself: can others bypass this company? If not, it’s a good tax collector. In summary, for stocks focus on—energy (nuclear) + shovel sellers (power grids, hardware, memory chips, etc.) For coins focus on—underlying settlement networks (BTC/SOL/PYTH/LINK/TAO) + stablecoin issuers The lesson from the past decade is clear: the application layer dies off repeatedly, while the tax-collecting layer remains steady. In the next decade, don’t be the chump; be the toll collector. DYOR 🙏