Panchaea Ltd

Following on from our recent overview of Biden's AI Diffusion Act, what are the biggest upsides and risk points? Risk-wise, the US is likely to face... 💵 Higher costs - chips made domestically are often more expensive, which could trickle down to AI businesses and end users. ⏰ Delays in relief - the benefits of increased production won’t be immediate, leaving current chip shortages unresolved for the short term. 👀 Global market reactions - encouraging local production could shift global dynamics, potentially leading to trade conflicts or restricted access to international innovations. What are your thoughts on how this will shape AI’s trajectory? #AI #Semiconductors #DiffusionAct #TechPolicy

The Diffusion Act from the Biden administration is set to have a substantial impact on the AI market 👇 At its core, the Act focuses on securing domestic chip production and reducing reliance on international supply chains, leaning on Know-Your-Customer (KYC) requirements for the transfer of new AI chips. So, why does this matter? 🤔 From training cutting-edge AI models to running real-time applications, the demand for GPUs and specialised silicon is skyrocketing. The surface level aim of this ruling is to incentivise chip manufacturing locally. The US aims to resolve delays that have previously stifled innovation. Biden's Act has been met with a mixed reception. OpenAI have championed the new policy, whilst NVIDIA's Vice President issued a less glowing statement on it here: https://lnkd.in/eZJHzUJV #AIDiffusionAct #AI #GPU

We had a great time meeting with our partners at Creturner Group AB (Publ) last week at NVIDIA's London offices 👏 We've got a big few months ahead with Cloud & AI Infrastructure London around the corner - watch this space for more news soon.

The view from NVIDIA's London office isn't half bad...IF the UK weather holds out. Given the utter insanity of Jensen's CES keynote, and the product roadmap for the next few years, it's tough trying to keep up with everything. And that's as someone working in the industry. With the benefit of having some understanding of what's going on. I can only imagine how difficult it must be to wrap one's head around it all from the outside. Especially with newer developments and technologies like those in the Agentic AI space. So, when you get the opportunity to hear what's next, straight from the proverbial green horse's mouth, you take it. And when you get to do that with Panchaea Ltd long term partners like Creturner Group AB (Publ), even better. Great to see Daniel Moström, and Johan Nyrén in person with Pete Overell after so many months of virtual calls. Also a huge thanks to Jim Hughes and Urvashi Mirpuri for hosting us and walking us through the NIM and Agent Blueprint ecosystem. Exciting times ahead for agents with Nvidia's new partnerships with CrewAI, LangChain, and the rest. And if CES was the benchmark, who knows what NVIDIA GTC will bring. It's going to be a big year. What was your biggest takeaway from Jensen's CES keynote? What are you most excited about? Let me know below!

The GPU seems to be taking on a life of its own. Growth is faster than I expected, which tells me it’s time it had a place to live. That’s why I've created a LinkedIn page. Here’s what you’ll get: • Updates on GPU charges and market trends • Energy challenges and how data centres are adapting • New developments in GPUaaS, NeoClouds, and AI infrastructure • A connection point for anyone serious about the future of compute and AI And for me? It’s a way to expand the conversation and continue LARPing as a LinkedInfluencer. If you’re serious about keeping up with the insights shaping compute, The GPU page is the best place to start. Oh, and if you’re not already subscribed to the newsletter? That’s where the real depth is. Let’s fix that too.

Falling prices, fierce competition, and the rise of AIaaS have sealed GPUaaS’s fate. 2025 will bury the old model for good.  Read more here: https://lnkd.in/eQQ_D5gg Follow for more timely news, views, and interviews. You may also subscribe free to access all content here: https://buff.ly/3XiMLeI At The Tech Capital you lead, we report. #TTC #TheTechCapital #YouLeadWeReport #AIaaS #ArtificialIntelligence #DigitalInfrastructure #GPU #GPUaaS #NVIDIA #Panchea

"Here lies GPU-as-a-Service (GPUaaS): a victim of its own success. "Once the hyper-profitable backbone of scalable infrastructure, it democratised access the high-performance hardware needed to train AI models without the upfront CapEx. But 2024 killed it." - Ben Baldieri Ben has published his first opinion piece on The Tech Capital, exploring the limits of GPUaaS and the next evolution into AI-as-a-Service. Read the full article to learn more: https://lnkd.in/eQQ_D5gg

How much could Google charge for access to the new Willow quantum chip on a chip/h basis if we take the NVIDIA H200 as a benchmark? Let's run the numbers! Sundar Pichai stated that Willow took 5 minutes to solve a problem that would have taken the most powerful supercomputers 10 septillion years. The most powerful supercomputer is Lawrence Livermore National Laboratory's El Capitan. This is enough data to work out a speed-up factor (S). First, let's convert the classical computation time into seconds, starting with seconds/year. 1y = 365.25 days × 24h/day × 3600s/hour 1y = 31,557,600 seconds This means the total number of seconds in 10 septillion years is: = 31,557,600 seconds × 10²⁵ = 3.15576 × 10³² seconds S is a function of Classical Computation Time / Quantum Computation Time: = 3.15576 × 10³² seconds / 300 seconds This means Willow is 1.05192 × 10³⁰ faster than El Capitan. Next, let's determine how many H200s are equivalent to El Capitan. We can do that using performance metrics. For simplicity, we're converting El Capitan's Rmax to FLOPs. But supercomputers are not the same as a single GPU. And El Capitan was built with AMD's MI300A. So it's not a perfect comparison. However, it does provide a context many will understand. And a GPU/h chargeability benchmark. So let's do it anyway: • El Capitan's Rmax: 1.742 exaFLOPs/s = 1.742 × 10¹⁸ FLOPs/s • Nvidia H200's FP8 Performance: 3,958 TFLOPs = 3.958 × 10¹² FLOPs/s The number of H200s is, therefore: = El Capitan Performance / H200 Performance = 1.742 × 10¹⁸ / 3.958 × 10¹² Which is roughly 439,800 H200s. Now, let's combine these calculations for Willow's equivalent number of H200s: = Number of H200 GPUs for El Capitan × Speed Up Factor (S) = 439,800 × (1.05 × 10³⁰) This means Willow's performance is equivalent to ~4.607 × 10³⁵ H200s. From here, we can estimate a HIGHLY SPECULATIVE hourly cost if we take an assumed $2.40/GPU/h per H200: Willow’s Hourly Cost = Willow’s Equivalent H200 GPUs × H200 Hourly Cost = 4.607 × 10³⁵ × $2.40 This gives us an hourly cost of ~$1.106 × 10³⁶. That's a big number. Let's put it into context. Global GDP in 2023 was ~$120T, or $1.2 × 10¹⁴. This means Willow's speculative hourly cost is: ~9.22 × 10²¹ times global GDP. Crazy. So, there you have it: Google’s Willow chip could hypothetically cost $1.1 × 10³⁶ per hour—roughly 9.22 sextillion times global GDP. Of course, this calculation is held together by assumptions, a pinch of salt × 10¹⁴, and sheer curiosity. Quantum breaks everything we think we know about everything. From pricing and performance to the FUNDAMENTAL NATURE OF REALITY (hello multiverse). The purpose of this thought experiment was to demonstrate this. But here’s the real question: If Google Cloud was charging a hypothetical $1.1 × 10³⁶/Willow/h, how much more expensive would Amazon Web Services (AWS) be? Let me know below!

If you're a GPUaaS provider, and 2024 has been a rough year for you, I have bad news. It's going to get worse. The market’s shifted. Margins are gone. GPUs are a commodity. Renting hardware alone just doesn’t cut it anymore. Enterprises need more than just compute - they need outcomes. Here’s what’s driving the shift: 1. Cloud repatriation: Public clouds like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud are expensive. Costs keep spiralling, and billing is opaque. That's a feature, not a bug. There's significant appetite (if my inbox is anything to go by) for bringing workloads back to private and hybrid clouds. AI is still a blue ocean, IMO. AIaaS providers identifying and dominating a niche are perfectly positioned to offer tailored, cost-effective solutions to satisfy these repatriation desires. 2. Data sovereignty: The EU AI Act and similar regulations make compliance non-negotiable. Sovereign clouds are, therefore, not just desirable but essential. Neoclouds are typically enough in their development to meet this need. 3. The talent gap: Enterprises don’t have the in-house expertise to build and manage end-to-end AI solutions. If they do, and the in-house talent is actually talented, the big tech companies routinely poach these resources with much better salary packages. I've heard stories of this happening several times in London. By bundling compute, storage, data analysis tools, fine-tuning, and more, AIaaS can solve this problem. Lower the barriers to entry, and the need for top-tier data scientists will reduce. 4. Agentic AI: We all know that LLMs on their own aren't the most useful tools. It's the same problem as offering just GPUs. Enterprise value lies in what's possible with those tools. This means agentic AI is the most exciting option because EVERYTHING is possible. People and processes define an organisation, and agentic AI impacts both. AIaaS platforms that understand this and make it easy for an enterprise to deploy agents at scale will win. 5. A shrinking pool of conventional GPUaaS clients: The biggest GPUaaS contracts tend to be the model builders. While they'll pay a premium for the latest generation of hardware, they increasingly want MUCH larger clusters. Few GPU providers have the capacity or credit-worthiness to support these deployments. Unless you have several hundred million $ in liquidity, best of luck to you. -- The writing is on the wall. GPU providers must pivot to offering solutions or face irrelevance. 2025 will reward adaptation. Providers who embrace AIaaS and deliver outcomes, not just hardware, will lead the next wave of innovation. Those who don’t will be left thinking about what could have been. However, hope is far from lost. Opportunity is abundant out there. You just need a little creativity to uncover it. DM Pete Overell or me, and we can show you where to dig.

Data centres are stuck in the past. And it’s costing them millions. The traditional DC PropCo-OpCo model has worked for years: • PropCo owns the real estate • OpCo handles the operations Margins come from leasing racks, either through: • Multi-tenant colo • Single-tenant wholesale But that model is slow. And in a world increasingly driven by AI and GPU compute, slow tends to be a target for disruption. Forward-thinking operators understand this. They've seen the writing on the wall for their old business model. They've realised that in a market where speed is everything, one practice makes sense above all others: Vertical integration. No longer just landlords, they’re deploying: • Storage • Compute • Networking • Desalination • Carbon offset • Waste heat recycling • Onsite power generation They’re becoming full-stack service providers. Why this shift? Two big reasons: 1. Faster ROI: Owing the infrastructure unlocks new, higher-margin revenue streams. Add revenues from onsite power generation, trading carbon offsets, selling waste heat, and the returns stack up much faster than traditional leasing. 2. Greater control and efficiency: Owning the entire stack (compute, storage, networking, power, and carbon offsets) means better optimisation, lower costs, higher performance, and stronger sustainability results. It makes sense for data centres to lead this transformation. They already have tenants hungry for compute and storage. They’re well-backed, with balance sheets to fund the required CapEx or debt. But here’s the catch: • Most data centres in Europe aren’t built for high-density workloads • Their infrastructure isn't suitable for GPU-heavy clusters That’s where modular deployments come in. Modular solutions let operators add high-density capacity quickly, without reworking their existing sites. Fast. Scalable. Perfect for the AI and GPU compute boom. This is where Panchaea Ltd can help. We’ve built relationships with multiple modular DC developers. We understand the complexities of designing, sourcing, and deploying GPU clusters. We know that success is about mastering: • Procurement and logistics • Cluster and modular DC design • Market force-resistant revenue models Helping you to: • Lower your cost of capital • Accelerate your route to market And giving data centre owners options: • Sell the PropCo: As a high-value, fully optimised asset • Sell the OpCo: To a hyperscaler or AI platform provider • Scale the services: Compete directly in the GPU compute and AI infrastructure markets The future of data centres is vertical integration. But it’s only an opportunity if you can execute it right. If you're ready to stop leaving margin on the table, let's talk. DM me or Pete Overell, and let's ensure your business model aligns with the future, not the past. -- Image taken from McKinsey & Company - link to the report in the comments!