Where Things Get Stranger: The Bleeding Edge

Opening

The previous three episodes mapped what exists. Hyperscalers, neoclouds, specialty silicon, databases, quantum, decentralized marketplaces, sovereign cloud build-outs. All of it commercial or imminently commercial. All of it something Pär could actually open an account with today if he wanted to. This bonus episode goes further out. The cloud world has a strange edge where the technology is real but the use cases are still science fiction, where commercial products exist but the buyers do not, where research teams ship prototypes that nobody will use for another five years. That is where we are going. Bleeding edge, not commercial edge.

The framing question for this episode is different. Instead of what should Pär know about, the question is what could Pär try. What can you actually spin up and play with today that uses compute primitives nobody else has access to. What is moving so fast that mid two thousand twenty six is a meaningfully different moment than mid two thousand twenty five was. Where is the next decade's default infrastructure being prototyped right now.

Eight territories. Thermodynamic computing. Wetware. Photonic processing. Off-Earth compute. Neuromorphic silicon. Fully homomorphic encryption. Million-year storage. And the things that are not public yet but are close enough to matter.

Thermodynamic computing

Every graphical processing unit on earth right now is doing something fundamentally absurd. It is using deterministic arithmetic to pretend to be probabilistic. Every diffusion model, every Markov chain sampler, every Bayesian inference job is burning megawatts of carefully ordered electrical signals to generate output that is, at the algorithmic level, supposed to be randomness drawn from a probability distribution. The thermal noise on the chip itself, which the chip's designers spent enormous engineering effort suppressing, is exactly the kind of randomness those algorithms need. The chip is fighting physics, then simulating physics, then producing a result.

Extropic is the startup that decided to stop fighting. The company was founded in twenty twenty two by Guillaume Verdon, an ex-Google quantum researcher better known online as Beff Jezos, with a team drawn from Google, IBM, Apple, and Microsoft. The product is the Thermodynamic Sampling Unit, abbreviated T S U. The architecture is built around what Extropic calls probabilistic bits, or p-bits. A p-bit is a circuit that natively outputs random samples from a programmable probability distribution, using thermal noise on the silicon as the source of randomness rather than as something to suppress. The energy cost of a single p-bit operation is about ten thousand times lower than the energy cost of a floating point addition on a standard graphical processing unit.

The hardware roadmap looks like this. X zero, the proof-of-concept silicon, shipped in the first quarter of twenty twenty five with a few dozen probabilistic circuits, enough to validate the physics. X T R zero, the development platform, shipped in the third quarter of twenty twenty five to research partners. It combines a traditional central processing unit with a field programmable gate array and two sockets for Extropic chips. Z one, the first production chip, is in early access throughout twenty twenty six. Each Z one contains roughly two hundred fifty thousand interconnected probabilistic circuits per chip, with millions per card, manufactured using standard complementary metal oxide semiconductor processes on the same fabrication lines that produce conventional silicon.

What you can do with this depends on what you are trying to compute. Thermodynamic computing is not a general purpose replacement for graphical processing units. It is specifically suited to probabilistic workloads. Diffusion model sampling. Bayesian inference. Monte Carlo simulation. Combinatorial optimization. Energy-based models. The kinds of problems where the answer is intrinsically a probability distribution rather than a single deterministic output. Extropic also released an open source Python library called T H R M L for developing and simulating thermodynamic algorithms on conventional hardware before they run on actual T S U chips.

The competitor is Normal Computing, which builds a similar category of probabilistic processing units using a slightly different architectural approach. Normal Computing is less public about its roadmap and has a more enterprise-oriented sales motion. Together, Extropic and Normal represent the most credible attempt to date at replacing a portion of the graphical processing unit workload with hardware that is fundamentally rethought from the silicon up.

For Pär, this is interesting because of the diffusion model angle. Pär's F L U X based LoRA training and inference workflows are diffusion model workloads. If Extropic's claims hold up at production scale, the same image generation that currently runs on RunPod H one hundreds could eventually run on T S U cards drawing one ten-thousandth the power. That is not happening this year, and probably not next, but the hardware exists today and early access partners are running real workloads.

Wetware

Now the genuinely strange section.

Two companies are currently selling commercial access to living human neurons as a compute substrate. The technology is called wetware, the science is called Synthetic Biological Intelligence, and the products are real things that you can pay for and access through a Python application programming interface today.

Cortical Labs is the Australian company that launched the C L one in March two thousand twenty five. The C L one is a desktop device, priced at thirty five thousand United States dollars, that grows real human neurons from stem cells directly onto a silicon chip. The chip reads electrical signals from the neurons in real time and responds with stimulation, creating a closed loop where the living neural network learns to perform tasks defined by the operator. The neurons are kept alive with fluid, nutrients, and temperature management. A new paper published this year introduced the C L application programming interface, which is a Python interface that lets researchers send signals into the neurons and read responses with sub-millisecond latency.

The cloud component is the Cortical Cloud, where remote access to C L one units costs about three hundred dollars per week. By early two thousand twenty six, the Cortical Cloud was running multiple server stacks, each holding up to thirty C L one modules. The first pilot data center, in Melbourne, was scheduled for fifty wet racks with sixty four bioprocessor modules each. The second, in Singapore, is targeting commercial alpha status by late two thousand twenty six. Both facilities are designed to Tier Three equivalent redundancy with biocontainment measures layered into the heating, ventilation, and air conditioning systems.

FinalSpark is the Swiss competitor, headquartered in Vevey, founded in two thousand fourteen by Fred Jordan and Martin Kutter. The Neuroplatform offers cloud access to brain organoids, which are three-dimensional clusters of brain cells about half a millimeter across, each containing roughly ten thousand living neurons. Organoids are grown from human stem cells and connected to multi-electrode arrays. The Python application programming interface lets external researchers stimulate the organoids electrically and measure their responses. FinalSpark uses neurotransmitters like dopamine and serotonin as chemical rewards to train the organoids, which is a fundamentally different learning mechanism than the gradient descent that underlies digital neural networks.

What you can actually compute on this hardware is the open research question. Pattern recognition, reservoir computing, control loops for robotics, and time-series prediction have all been demonstrated. Whether any of this scales to compete with digital neural networks on production workloads is unknown, and the consensus in the research community is that scaling to anything like the complexity of a transformer is many years away, possibly decades. The human brain runs on about twenty watts. Current frontier model training runs consume tens of megawatts. The structural energy gap is real, and biological substrates are the only known compute medium that achieves brain-like efficiency, but reaching commercial viability for general purpose workloads requires breakthroughs that have not happened yet.

The ethics are also genuinely unresolved. These are living human neurons, harvested from stem cells, kept alive in incubators, and trained through electrical and chemical stimulation. The companies involved publish position statements about sentience and moral consideration. The neurons themselves are far below any threshold of consciousness or self-awareness that anyone has seriously proposed, but the cultural conversation about wetware compute is just beginning.

For Pär, FinalSpark in particular is the kind of thing that is genuinely accessible. Three hundred dollars per week is real money but not absurd. The Python application programming interface makes integration straightforward. There is no practical use case for a Swedish operator running living neurons in Switzerland, but as an exploratory project that produces stories worth telling, this is one of the more interesting weeks a person could spend.

Photonic computing

The next category is light. Specifically, replacing some portion of conventional electronic computation with photonic operations performed in optical circuits using actual photons instead of electrons.

The dominant European player in this category is Q dot ANT, headquartered in Stuttgart, founded in two thousand eighteen. The product is the Native Processing Server, which Q dot ANT brought to commercial availability in late twenty twenty four. The architecture is built on thin-film lithium niobate, abbreviated T F L N, which is a material with high electro-optic coefficient, low optical loss, and thermal stability. Lithium niobate has been used in optical telecommunications for decades. Q dot ANT's contribution is figuring out how to manufacture it as integrated circuits and how to design computational primitives that compose into useful operations.

The performance claim is up to thirty times the energy efficiency and fifty times the performance of conventional processors for specific artificial intelligence and high-performance computing workloads, particularly nonlinear math operations like Fourier transforms. A Fourier transform that requires millions of transistors to compute digitally can be accomplished with a single optical element. The Native Processing Server plugs into existing data center hardware via standard P C I express interfaces and operates as a co-processor alongside central processing units and graphical processing units, not as a replacement.

Q dot ANT raised a sixty two million euro Series A in July two thousand twenty five, which made it one of the larger European deep tech rounds of the year. The second generation N P U was unveiled at Supercomputing two thousand twenty five in November. In April this year, Q dot ANT opened its United States headquarters in Austin, Texas, and hired Bruno Spruth, formerly the vice president of Power processor development at International Business Machines, as Chief Technology Officer. The expansion signals that commercial photonic computing is being treated as a real product category, not a research curiosity.

The United States competitor is Lightmatter, which has been pursuing a similar vision with a different chip architecture and has shipped its Envise product to selected customers. Lightelligence is another. Luminous Computing operated in the same category before pivoting. Ayar Labs is adjacent, focusing on optical interconnects between chips rather than optical compute within chips, and NVIDIA is among its investors. The thesis across all of these companies is the same. Light moves faster than electrons, generates almost no heat in linear operations, and can perform certain mathematical functions in a single physical step that would require thousands or millions of transistor operations to compute digitally. The catch is that not all computations are amenable to optical implementation, the manufacturing processes are different from standard complementary metal oxide semiconductor, and the software toolchains are immature.

For Pär, photonic computing is academically interesting but not yet relevant. Q dot ANT's Native Processing Server is a data center product priced for enterprise procurement, not a card you order from a website. The relevant moment for individuals will be when photonic accelerators become available through neocloud platforms or hyperscaler instance types, which has not happened yet but is likely within two to three years.

Off-Earth compute

This is the section where the strange becomes literal.

Two categories of off-Earth data center are now being built and operated commercially. Orbital data centers in low Earth orbit, and lunar data centers on or around the Moon. The technical case for both rests on three structural advantages. Free solar power, twenty four hours per day, with no atmospheric absorption losses. Free radiative cooling directly into the vacuum of space, with no need for water-cooling infrastructure. And a regulatory environment governed by space law rather than terrestrial jurisdictions, which is interesting for sovereignty-sensitive workloads.

Lonestar Data Holdings is the leading lunar player. The company began operations in twenty twenty three with a software-only payload aboard the Intuitive Machines I M one mission. In February twenty twenty four, Lonestar launched its first physical hardware to lunar orbit. In March this year, the company's Freedom payload landed on the Moon aboard Intuitive Machines Athena lander, carrying an eight terabyte solid state drive and a Microchip PolaFire system-on-chip field programmable gate array. The payload successfully completed file upload, download, encryption, decryption, and edge processing tests during the cislunar transit and on the lunar surface, including data handling for customers like Valkyrie A I and the Exploration Institute.

The commercial product is called StarVault, billed as the world's first commercially operational space-based sovereign data storage platform. Lonestar is taking capacity reservations for the first StarVault launch in October twenty twenty six. The longer roadmap is six purpose-built lunar storage spacecraft launching between twenty twenty seven and twenty thirty, each carrying multi-petabyte storage and edge processing capability, positioned at the Earth Moon L one Lagrange point. The eventual goal is to install hardware in lunar lava tubes, where the underground temperature is roughly negative twenty degrees Celsius and stable year round, providing both cooling and radiation shielding.

Starcloud, formerly known as Lumen Orbit, is the orbital data center play. The company has raised more than twenty million dollars, partnered with NVIDIA's Inception program, and in December twenty twenty five successfully trained a Google Gemma model in orbit using commercial-grade NVIDIA H one hundred graphical processing units. The training run was of a small Nano G P T model on the complete works of Shakespeare, which is a toy workload, but the demonstration was the first definitive proof that production-grade artificial intelligence training can run on space-based hardware. Starcloud's longer term plan is to operate as an energy and compute provider that hosts whatever chip architecture customers want to deploy, similar to how Crusoe positions itself terrestrially. Starcloud has named Crusoe Cloud as a strategic partner.

The other names worth knowing in this space are Axiom Space, which is collaborating with Spacebilt on optically interconnected data center infrastructure for the International Space Station scheduled for twenty twenty seven, K Two Space, which raised a two hundred fifty million dollar round for satellite platform development, Aetherflux, which describes itself as building an American power grid in space with applications in orbital artificial intelligence compute, and Google's Project Suncatcher, which plans to launch two prototype satellites carrying Tensor Processing Unit chips in early twenty twenty seven. SpaceX is also exploring using Starlink Gen Three satellites for data processing, and Elon Musk is reportedly seeking billions in funding for that initiative.

For Pär, off-Earth compute is purely a curiosity at this stage. The relevant detail is that the technical viability has been demonstrated in twenty twenty six, which means that within the timeframe of his current businesses, orbital data centers will move from theoretical to procurable. The cost economics already favor space for certain workloads. According to Lonestar's chief executive, terrestrial facilities operate at about five cents per kilowatt hour for the cheapest power, while a celestial data center can theoretically reach about one tenth of a cent per kilowatt hour including launch costs. Whether that math holds in practice will be tested over the next three to five years.

Neuromorphic silicon

A different attempt at rethinking the chip from the ground up is neuromorphic computing. The pitch is event-driven computation that mimics the spiking behavior of biological neurons rather than the clocked synchronous computation of conventional digital chips. Information is carried in the timing and frequency of discrete events rather than in continuous voltage levels. The energy efficiency advantage is supposed to come from the fact that idle parts of the chip consume essentially no power, because no events are occurring.

Intel's Loihi Two is the most widely accessible research neuromorphic chip, available through Intel's neuromorphic research program. The chip contains roughly one million artificial neurons and supports complex spiking neural network workloads with sub-microsecond inference latency on specific problem classes. International Business Machines NorthPole, announced in late twenty twenty three, demonstrated record-setting energy efficiency on standard image classification benchmarks, drawing about twenty five times less power than conventional graphical processing units on equivalent inference workloads.

BrainChip Akida is the commercial player most aggressively pursuing edge deployment. The Akida chip is a neuromorphic accelerator designed for always-on artificial intelligence in low-power devices like sensors, wearables, and embedded systems. SynSense Speck targets a similar market with event-based vision processing for cameras and robotics. Innatera and Prophesee are other commercial entrants. None of these chips are accessible through cloud platforms in the way that conventional graphical processing units are. Access is generally through development kits, research programs, or direct customer engagements.

The honest assessment is that neuromorphic computing has been roughly five years away from broad commercial relevance for about twenty years. The energy efficiency claims are real on specific workloads but the software toolchains are immature, the programming models are unfamiliar to most engineers, and the workloads where neuromorphic chips dramatically outperform graphical processing units are narrower than the marketing suggests. For Pär, this is interesting to know about as a future direction but not actionable in any near-term sense.

Fully homomorphic encryption

Fully homomorphic encryption, abbreviated F H E, is a cryptographic technique that lets you perform computation on encrypted data without ever decrypting it. The input is encrypted, the computation happens on the encrypted values, and the output is encrypted in such a way that the original key holder can decrypt it to get the correct result. Anyone watching the computation, including the cloud provider running it, sees only encrypted data throughout the entire process. This solves the structural privacy problem of cloud computing, which is that hosting your data on someone else's hardware fundamentally requires trusting that someone else not to look at it.

F H E was proven mathematically possible in two thousand nine by Craig Gentry. For about a decade after that, it was too slow to use for anything practical. A simple encrypted operation took thousands of times longer than the equivalent unencrypted operation. The recent breakthroughs are in scheme design and hardware acceleration that have brought F H E from impractical to merely expensive.

The leading commercial player is Zama, a Paris-based open source cryptography company. Zama publishes the T F H E r s library, which is a Rust implementation of the Torus F H E scheme optimized for boolean and integer arithmetic, and Concrete M L, which is a Python framework for running machine learning models on encrypted data. Zama recently launched its blockchain protocol with token sale, which is the cross-chain confidentiality layer that lets smart contracts run on encrypted state across Ethereum and Solana. Ledger, the hardware wallet company, is among Zama's institutional partners.

The competitors are Duality, IBM HE lib, Inpher, and a long tail of academic open source projects. The hardware angle is interesting too. Niobium and Optalysys are building F H E specific hardware accelerators that promise orders of magnitude speedup over running F H E on conventional graphical processing units.

For Pär, F H E is one of the genuinely promising bleeding edge technologies. The Pärception consulting practice will eventually encounter clients with workloads that touch regulated data, where the inability to decrypt the data on a cloud provider's hardware is a structural obstacle. F H E is the cryptographic answer to that problem. The practical answer today is usually a Trusted Execution Environment like Intel T D X or NVIDIA confidential computing on H one hundreds. The F H E answer becomes practical when hardware acceleration brings the cost premium below one order of magnitude, which several roadmaps are pointing at within the next two years.

Million-year storage

This category is small but conceptually interesting. The premise is that conventional storage media degrade over years or decades. Magnetic tape, the cheapest archival medium, lasts roughly thirty years before requiring re-archiving. Solid state drives and hard drives have shorter lifetimes. For genuinely long-term storage, where the data should outlive the engineers who created it, conventional media are not adequate.

Microsoft Project Silica encodes data into fused silica glass plates using femtosecond lasers, which carve nanoscale voxels into the glass that can be read back optically. The estimated lifetime is ten thousand years. The storage density is around seven terabytes per glass plate the size of a coaster. Project Silica is still a research project at Microsoft Azure with no commercial availability date announced, but Azure has been running pilot programs with selected enterprise customers since twenty twenty four.

Cerabyte is the German commercial competitor in the same category. The Cerabyte system writes data onto ceramic-coated glass plates using laser writes, with claimed durability of over five thousand years and capacities expanding to petabyte scale per rack unit. Cerabyte has shipped prototype systems to enterprise customers and is targeting full commercial availability over the next two years.

The synthetic D N A storage category is the more exotic competitor. Companies like Twist Bioscience and Catalog Technologies encode digital data as nucleotide sequences in synthesized D N A strands. Storage density is theoretically extreme. A single gram of D N A can hold roughly two hundred fifteen petabytes of data. Read and write times are slow and the cost per terabyte is still high, but the durability under proper preservation conditions is essentially indefinite. D N A is how biology has stored information for four billion years, and the molecular stability is well understood.

For Pär, million-year storage is academically interesting but currently irrelevant. The Pärkit historical telemetry will not still be useful in ten thousand years, and the cost per terabyte today is several orders of magnitude above conventional cloud storage. The relevant moment for million-year storage is when an archive emerges that genuinely needs to outlive civilization. Cultural heritage projects, scientific record archives, and long-duration space missions are the obvious early adopters.

What is not public yet

The final category is the rumored, the leaked, and the imminent. Nothing in this section is confirmed public product, but the signals are credible enough to mention.

Google's Project Suncatcher launches in early twenty twenty seven with two prototype satellites carrying Tensor Processing Units. The longer plan is for full orbital training clusters built on optical interconnects between satellites, which would be the first hyperscaler-grade artificial intelligence training facility outside Earth's gravity well. Google has not published a commercial roadmap beyond the prototype phase.

Anthropic and Open A I have both alluded to confidential computing offerings using NVIDIA H one hundred confidential mode, where the model weights and the customer's data are both encrypted in graphical processing unit memory and cannot be inspected by Anthropic or Open A I employees. The technology exists today. The product packaging for general availability has not yet shipped.

The European Union artificial intelligence gigafactories are tracking through the procurement process during twenty twenty six, with the first major contracts expected to be announced in the second half of the year. The Mistral Swedish data center construction is underway, with operational milestones expected within twelve to eighteen months.

The Cerebras initial public offering, which priced last week as we recorded the previous episode, has set a benchmark for what specialty silicon companies can achieve as standalone public companies in the post-Groq-acquisition environment. PsiQuantum is expected to follow with its own public offering during twenty twenty six. Quantinuum has filed an S one and is awaiting market conditions.

NVIDIA is rumored to be preparing the Blackwell Ultra refresh for late twenty twenty six, with the Vera Rubin architecture following in twenty twenty seven. The Rubin generation is expected to integrate the absorbed Groq techniques as first-class capabilities, which will further raise the bar for any remaining specialty silicon competitors. The competitive question for the next two years is whether Cerebras can defend its wafer-scale advantage against successive NVIDIA generations, and whether any new specialty silicon entrants emerge.

Closing

This bonus episode is the strangest of the series and has the least practical application. Almost nothing covered here is going to change anything Pär does this year, or possibly next. The Extropic Z one is in early access. The Cortical Cloud and FinalSpark Neuroplatform are accessible but produce experimental results rather than production outputs. Q dot ANT photonic processors are not yet available outside enterprise data center contracts. Lonestar's commercial StarVault launches in October but is positioned as premium disaster recovery, not general compute. F H E is workable but slow. Million-year storage is irrelevant on any timescale that matters to a working operator in twenty twenty six.

The value of paying attention to this layer is different. The technologies that look weird and useless right now are the ones that define the default infrastructure of the twenty thirties. The graphical processing unit looked weird and useless in two thousand seven. The mobile phone looked weird and useless in two thousand. The internet looked weird and useless in nineteen ninety three. The pattern repeats. What is being prototyped on the bleeding edge today is what becomes the standard procurement decision a decade from now.

If Pär wanted to spend a weekend on one of these, FinalSpark is the most viscerally strange. Three hundred dollars per week buys real-time Python application programming interface access to living human neurons growing in a Swiss laboratory. There is no practical use case. There is no business outcome. There is just the genuine experience of writing code that talks to neurons that are alive while they receive the signal. Most operators will not have done this by twenty thirty. Doing it in twenty twenty six is the kind of thing that produces stories, that informs the consulting practice in indirect ways, that maintains the reflex of treating compute as a genuinely strange and varied substrate rather than a uniform commodity.

That is the bleeding edge as of mid May two thousand twenty six. The four-episode series ends here. The map is now complete enough to be useful, and incomplete enough to be honest about what is changing fastest. The next episode of this feed will be about something else.