Topusko AI Data Center – One Gigawatt of Silicon and the Future of Computing
Inside the proposed 1 GW campus, the five questions a gigawatt AI factory has to answer, and the photonic, analog, fluidic, chemical, neuromorphic and quantum machines that might one day make a building this large unnecessary.
The provided text examines the severe energy and thermal bottlenecks facing modern artificial intelligence infrastructure, particularly driven by massive gigawatt-scale silicon data centers like the planned Pantheon facility in Croatia. It explores various post-silicon computing paradigms—including photonic, analog in-memory, neuromorphic, fluidic, chemical, and quantum architectures—that offer drastically lower energy consumption by processing information natively outside traditional electronic frameworks. Despite their theoretical efficiencies, these alternative substrates face significant obstacles such as precision scaling limits and software ecosystem inertia, notably the dominance of NVIDIA's CUDA platform. Consequently, while the industry continues to deploy brute-force complementary metal-oxide-semiconductor (CMOS) hardware to meet immediate demands, it is increasingly forced to consider hybrid architectures that combine digital processing with optical transport and analog calculation.
▶️ Rave the World Radio
24/7 electronic music streaming from around the globe
Power file · Topusko · Pantheon AI
They want 1 gigawatt of hot silicon in Topusko. The other computers are not ready.
A spa town. A gigawatt hall. Read the bill.
Pantheon Atlas wants a NVIDIA-style AI factory on 310 acres in Croatia. Light, analog memory, spikes, bubbles, chemistry and qubits could burn less power. None of them can take this hall’s job in 2027.
The hall
1 GW / 800 MW IT
The bill
€50bn across the life · €12bn to start
The industry’s answer to bigger models is a bigger building. Pantheon Atlas LLC has proposed the Pantheon AI campus for Topusko, designed to NVIDIA gigawatt-scale AI factory standards. The report calls it one of the largest private United States investments planned in Europe, and the largest investment figure ever attached to a project in Croatia.
The campus is drawn for 1,000 MW in total and 800 MW of usable IT load, on 310 acres that can grow to 450. Construction is aimed at the second quarter of 2027. Operations are aimed at the first quarter of 2029. Capital in the report runs to €12 billion to put the foundation in, and up to €50 billion ($58 billion) across the life of the site. Europe’s established hubs, the same report says, are running with vacancy under eight percent and long waits for a grid connection. Topusko is the proposed way around that queue.
One campus, run as one supercomputer. The report’s price for staying on ordinary silicon is €12 billion up front and up to €50 billion over the life of the site.
What you need to know
1 GW / 800 MW — total capacity, and the share reserved for computers.
€12bn, then up to €50bn — opening works, then the life of the campus.
€500 million — grid works: 280 km of new 400 kV lines and a Topusko substation.
500 MW solar, 8,000 MWh batteries — on-site generation and storage in the plan.
Not yet — photonic, analog, neuromorphic, fluidic, chemical and quantum chips do not replace this hall today.
The wires and the water
Pantheon Atlas has a supply blueprint approved by HOPS, Croatia’s transmission operator. The upfront grid bill in the report is €500 million: 280 kilometres of new 400 kV lines and a 400 kV substation in Topusko. When the lines are done, the report says they pass to the Croatian state, with about €80 million a year in network fees for HOPS.
On the campus itself the plan adds a 500 MW solar plant and an 8,000 MWh battery. The same blueprint is described as unlocking 5.2 GW of new clean capacity on the national grid. A continuous gigawatt is about 8,000 GWh a year.
On the campus · 310 acres
Expandable to 450. Target open date: first quarter of 2029. Construction start in the plan: second quarter of 2027.
On the grid · 400 kV
New substation in Topusko, 280 km of new lines, €500 million up front. The report says the lines are then handed to the state.
Why the project wants 1 GW
Frontier models hold trillions of parameters, and the training step wants them updated together. Under NVIDIA’s DSX reference design, a gigawatt AI factory is not a warehouse of independent servers. It is drawn as one supercomputer.
The clock they are buying is time-to-first-token: how fast the building starts returning intelligence after the capital is spent. Spread the same gigawatt across ten 100 MW sites and the chips have to talk across a wide-area network. NVLink is built for microseconds between nodes in one fabric. If the data is late, the processors sit idle, and watts-per-token falls apart.
A single hall also changes the power plumbing. An 800 volt DC architecture, taken straight from the medium-voltage grid, skips repeated conversions from AC to DC. The report’s strategic claim is simple: Europe does not yet have one room dense enough for this class of job, so Topusko is drawn for the hyperscalers who want that room.
Which jobs actually need the GPUs
A website does not. A database does not. Ordinary x86 CPUs still carry that work. The GPU hall is for dense matrix math.
First on the list is training foundation models and large language models. Backpropagation computes a storm of gradients at once, which is what thousands of GPU threads are for. Next are agent sandboxes, where systems stay awake in many kinds of input and keep working a problem. Then the physical jobs: computational fluid dynamics, molecular dynamics for drug discovery, and digital twins rendered on something like NVIDIA Omniverse. Tensor cores do the floating-point work. For that list, a CPU is the wrong tool.
Could other chips cut the power?
On paper, yes. Photonic circuits compute with light, so they skip the resistance and the waste heat of an electronic wire. The report cites projections of up to 30 times the energy efficiency and 50 times the speed of current digital GPUs. Analog in-memory computing attacks a different waste: the von Neumann commute, shuttling weights from memory to a processor and back. Do the multiply-accumulate inside the memory array, by Ohm’s law and Kirchhoff’s law, and the report’s range for specific AI operations is 10 to 100 times less energy. Neuromorphic chips copy the brain’s trick. They spike when a threshold is crossed, and they sit near idle when it is not.
Then the report says the quiet part. Precision is hard to scale. There is no CUDA for these parts. Technology readiness for general, high-precision training is low. They cannot replace a 1 GW general-purpose GPU hall today. The report calls them the plausible path of the next decade, which is a different sentence from “cancel the concrete.”
What can leave the big hall
Training and answering are different jobs. Training a foundation model has to keep every chip in step, because the gradients have to meet. That is why ten small buildings do not equal one big one.
Inference is the finished model answering a person. It does not need tens of thousands of chips locked together. That work can move to smaller sites, closer to the user, on local renewable microgrids. Analog in-memory chips and edge photonic chips fit that smaller job. They do not need a gigawatt roof.
What can be built later
Operators like to build in slices, because capital hates an empty hall. The grid does not work that way. A utility will not dig the landscape twice. For Topusko, the report says the 400 kV substation and the transmission lines go in for the full theoretical load at the start. The source describes that backbone as four dedicated lines.
Inside the fence, the computers can wait. A digital twin — the NVIDIA Omniverse DSX blueprint is the one named — lets engineers rehearse heat, electrical load and airflow before the racks arrive. The report’s phase runs from an opening 100 MW of IT toward the full 800 MW, in step with demand that has actually appeared, with the power spikes checked at millisecond scale so a staggered switch-on does not cook the room.
Light, analog, bubbles, chemistry, spikes, qubits
Light. Electrons in copper and silicon make heat, lose the signal over distance, and run out of pins. Photons do not shove each other aside, so many colours of light can share a path, and they dump almost no heat. Optical neural networks try to turn the GPU’s matrix step into a calculation on a photonic chip. One industry roadmap cited in the report runs photonic throughput from 0.1 GOps in 2024 toward 100,000 GOps by 2028.
The trap is the doorway. If every nonlinear step drops back into electrons, the lasers and the converters can eat the gain. Real optical computing has to do the kink — the ReLU, the sigmoid, the softmax — in light. The report describes thin-film lithium niobate Mach-Zehnder modulators used that way, including an optical stand-in called Optmoid and a relative called Optmax, tried on vision transformers and on a GPT-2 trained on FineWeb-Edu. It says CIFAR-10 accuracy tracked the digital softmax, and that the setup still worked in strong noise at encoding speeds up to 10 gigabaud.
The source left the measured accuracy, the bias term, the noise figure and the device size blank. Those blanks stay blank here. A separate line of work puts quantum emitters inside gallium phosphide nanostructures that saturate on a few photons. That paper’s estimate for large models is optical power below 1.2 W. It is a research path, not a procurement line for Topusko.
Analog memory. Since the 1940s the usual computer has kept the numbers in one place and the arithmetic in another. In an AI workload the commute costs more than the sum. Analog in-memory computing stores the weights as conductances in a crossbar — resistive RAM, phase-change memory, or capacitorless IGZO DRAM — and lets the currents add themselves in one step.
The bill comes due at the analog-to-digital converter. The currents are analog. The next layer wants bits. Converter power climbs hard as you ask for more precision, and slicing a precise number into many cheap conversions can wipe out the savings. The report points at two escapes. A hierarchical charge-domain design, Charge-CIM, is described as cutting ADC energy by 91.7 percent and lifting overall energy efficiency 2.7 times against the designs it was compared with. An analog shift-add converter is described as living with up to 7-bit quantization loss and reaching up to 37 times lower energy-delay product than a conventional flash ADC.
Bubbles. Fluidic logic is older than the microcomputer. The modern version, from the Prakash Lab, uses bubbles in microchannels as the bits. A bubble changes the resistance of the channel, and the next bubble takes a different door. The family includes AND, OR and NOT, flip-flops, counters and oscillators. The logic itself needs no electricity, and it does not care about electromagnetic interference, radiation or brutal heat. It will not train a language model. It can run a test where the sample and the computer are the same fluid.
Chemistry. Chemical computing goes further from the bit. Concentrations are the data. The reaction is the program. Reactions in the Belousov-Zhabotinsky family can settle into a shortest path or a spatial optimum by going to equilibrium. The report keeps this in the lab and in theory. It is not a training cluster.
Spikes. Neuromorphic chips run spiking networks. Power falls toward zero when nothing happens, which is why they suit an event camera more than a gigawatt training run. A spike is not a smooth curve, so the backpropagation used on GPUs does not drop onto these chips cleanly.
Qubits. Quantum computing adds qubits, and with them superposition and entanglement. The report also notes quantum nonlinear optics, where entangled photons are asked to do the kink that ordinary optical materials do weakly. Cryogenics, error correction and coherence keep that machine out of the near-term generative-AI factory.
Why the money still goes to silicon
If light can promise 30 times and analog memory can promise two orders of magnitude on the operations that fit it, the obvious question is why Croatia is being asked to host a brute-force GPU campus. The report’s answer is the hardware lottery, plus a software moat with a name: CUDA.
PyTorch, TensorFlow and the libraries around them are tuned to the memory bandwidth, the tensor cores and the floating-point habits of digital GPUs. Moving a gigawatt hall onto analog noise, quantization limits and continuous signals would mean rewriting that stack while rivals keep shipping. In a race measured in months, capital goes to the machine the software already knows: DSX, 800 volt DC, and a digital twin that sweats the last gains out of a substrate the report itself calls inefficient.
The bridge already on the table is co-packaged optics. Photonic transceivers move onto the silicon package and carry the fast input and output. Copper stops being the bottleneck for the commute. The arithmetic stays digital, so the software does not have to move. That is a smaller revolution than a new kind of computer, and it is the one a 2027 building can actually buy.
Where each machine stops
CMOS GPUs. Good at precision, and a software world that already runs. Stopped by heat, the memory commute, and power that scales with the work.
Analog memory. The math happens where the weights sit. 10×–100× on fitting operations. Stopped by converters, quantization noise, and analog drift.
Photonic. Almost no heat in the light path. Up to 30× in the projections cited. Stopped by converting back to electrons, and by the size of the optical parts.
Microfluidic. Logic with no electricity. Fine in radiation and interference. Slow. A specialist. Not an AI factory.
Hold the percentages apart
Electricity. The report says a continuous gigawatt is about 8,000 GWh a year, and calls that 9.5 percent of Croatia’s electricity. The note beside that sentence gives two 2024 baselines: 84,000 GWh of total energy, and 19,000 GWh of electricity. Nine and a half percent is 8,000 divided by 84,000. That is the total-energy baseline. Eight thousand divided by 19,000 is about 42 percent. Headlines in the works cited, including Poslovni.hr and Total Croatia News, say the hall would use electricity on the scale of a third of the country. Three readings. One building. They do not belong in the same sentence.
Important
Build the hall for the demand that exists. Do not describe a research chip as a substitute you can pour in 2027. Training stays together. Inference can leave. The power lines go in for the full load. The racks can wait for proof.
Topusko is a silicon factory because the software already runs there. The machines that sip power are real. They are not this building.
Figures follow the source report and the works it cites. The 11.4 million gallon water figure is carried there as an AI estimate. Optical accuracy, bias, noise and device size were blank in the source and are not filled in here. The 9.5 percent, the roughly 42 percent arithmetic on the report’s own electrical baseline, and the “one third of Croatia” headlines are different measurements.
References
1. Pantheon AI — Leading Europe Into The Future. (pantheonai.com)
2. AI Magazine — Pantheon AI’s US$58bn Data Centre Campus. (aimagazine.com)
3. Data Center Dynamics — Pantheon Atlas secures grid approval for 1GW Croatia data center. (datacenterdynamics.com)
4. Datacenterslist — Pantheon Atlas Topusko Data Center. (datacenterslist.com)
5. Balkan Green Energy News — Croatian Pantheon AI data center clears major grid hurdle. (balkangreenenergynews.com)
6. Poslovni.hr — Pantheon AI centar u Topuskom, headline on electricity “kao trećina Hrvatske”. (poslovni.hr)
7. Klimatski.hr — “Najveća investicija u povijesti Hrvatske”: PR-ovska AI magla, 30 April 2026. (klimatski.hr)
8. Total Croatia News — Pantheon Data Centre Will Consume as Much Electricity as a Third. (total-croatia-news.com)
9. Bug — Projekt Pantheon: 50 milijardi eura za podatkovni centar kod Topuskog. (bug.hr)
10. Tom’s Hardware — Nvidia reference design for gigawatt-scale Omniverse DSX data centers. (tomshardware.com)
11. NVIDIA — Omniverse DSX Blueprint. (docs.omniverse.nvidia.com)
12. Siemens — From chip to grid: the digital twin for NVIDIA DSX AI factories. (blog.siemens.com)
13. NVIDIA — AI Factories. (nvidia.com)
14. NVIDIA On-Demand — Behind gigawatt-scale AI factories, GTC Berlin. (nvidia.com)
15. NVIDIA — 800 VDC power architecture for the AI factory. (blogs.nvidia.com)
16. StorageNewsletter — Oracle AI World 2026: Vera Rubin at multi-gigawatt scale. (storagenewsletter.com)
17. Radiant — NVIDIA Omniverse DSX overview. (radiant.co)
18. Q.ANT — The photonic AI accelerator. (qant.com)
19. Lighting the Way for AI: Photonics in Computing as of Mid-2026. (reddit.com)
20. In-memory Computing Architectures for Energy-efficient AI. (researchgate.net)
21. Quantum Nonlinearity for Optical Neural Computing. (arxiv.org)
22. Enlit — Pantheon AI data centre to power itself with renewables in Croatia. (enlit.world)
23. Data Centre Magazine — Gigawatt AI factory blueprints with digital twins. (datacentremagazine.com)
24. Data Center Frontier — Britain’s bargain with the AI data center industry. (datacenterfrontier.com)
25. PatSnap — Optical neural networks and photonics. (patsnap.com)
26. Photonic-Electronic Integrated Circuits for High-Performance Computing. (arxiv.org)
27. Integrated electro-optic attention nonlinearities for transformers. (arxiv.org)
28. Same paper, earlier HTML version. (arxiv.org)
29. Analog In-Memory Computing Attention Mechanism. (arxiv.org)
30. Investigating Energy Bounds of Analog Compute-in-Memory. (arxiv.org)
31. End-to-End Analog Computing-in-Memory Architecture. (arxiv.org)
32. imec — Analog-to-Digital Converter Design Exploration for Compute-in-Memory. (imec-publications.be)
33. Manu Prakash — Microfluidic Bubble Logic, MIT thesis. (cba.mit.edu)
34. University of Washington ECE — Computing with Fluids. (ece.uw.edu)
35. Prakash Lab — Microfluidic Bubble Logic. (prakashlab.stanford.edu)
36. Pressure-Driven Two-Input 3D Microfluidic Logic Gates. (pmc.ncbi.nlm.nih.gov)
37. Non-Newtonian droplet-based microfluidics logic gates. (researchgate.net)
38. Large-scale photonic computing with nonlinear disordered media. (arxiv.org)
39. Energy bounds of analog compute-in-memory, PDF. (arxiv.org)
40. Compute Compass — Photonic computing companies in 2026. (computecompass.com)
Comments
Post a Comment