Talk to the founders
Talk to the foundersinvestors@synapsenet.io

AI guesses. In banking, defense and infrastructure, a guess can cost millions.

And a right answer is worthless if its data can be stolen, decrypted or seized. Synapse Net fixes both, with an AI that checks its answer before it speaks.

Talk to the foundersGet the investor deckGet the investor deckGet the investor deck
See the problem · Seed round open, $1M

Built on open standards and the hardware the seed round buys

Is this ChatGPT? No.

ChatGPT predicts the next word. PRISM checks the answer before it gives one.

Chatbotone word at a time, wrong at the endPRISMchecks the whole answer first
ChatGPT and other chatbots
  • Picks each next word because it is likely, not because it is true.
  • When the prediction is wrong, nothing inside it notices.
  • Ask it why, and it predicts an explanation too, even for a mistake.
PRISM
  • Checks the whole answer against what it knows before it speaks.
  • A contradiction makes the energy spike, so it corrects itself first.
  • Shows what happened while it reasoned, step by step.

Chatbots are built to sound right. PRISM is built to be right.

Why prediction is not reasoning

Three steps, before a single word.

Chatbots write first and never check. PRISM learns what is true, checks the whole answer, and corrects itself before anything reaches you.

time
cause
place
price
date
law
sum
size
Checking time

Learns what is true

PRISM learns basic truths and how concepts relate, instead of which word usually comes next.

PRISM
Hold the payment: the arrival date contradicts the departure.
Step 2 · InvoiceAxiom · time order
Step 2 · Supplier invoice · classified false“Arrived 1 March, two days before it left.”

Shows its reasoning

Every answer comes with the steps behind it, and what PRISM classified as true or false.

This answer128 checks
Customs date matches the logFits
Amount matches the contractFits
Arrival before departureEnergy spike
Correcting the thought
Currency is EURFits
Bank details unchangedFits
Supplier ID verifiedFits
Invoice number is uniqueFits

Corrects itself first

When a contradiction makes the energy spike, PRISM fixes the thought before a word reaches you.

Two problems, and both are expensive.

First, the AI itself can't be trusted. Second, even when it is right, the data it runs on can be stolen, decrypted or seized. Banks, grid operators and governments are told to adopt AI anyway.

95%

AI spending with no return.

MIT researchers found that about 95% of the organizations they studied saw no measurable profit impact from generative AI. A system that can be confidently wrong never gets trusted with the decisions that pay.

1 · The AI can't be trusted · MIT NANDA, The GenAI Divide, 2025
Dec2026

Wrong answers become your liability.

From 9 December 2026, EU product liability covers software and AI, and insurers have started excluding AI errors. If the AI can't show why it decided, the company can't defend the decision.

1 · The AI can't be trusted · Directive (EU) 2024/2853 · Verisk ISO exclusions, Jan 2026
$4.44M

The average cost of one data breach.

Globally, and $10.22 million in the US. For a government, stolen or corrupted data is measured in national security, not dollars.

2 · The data isn't safe · IBM, Cost of a Data Breach Report 2025
2030

Today's encryption has an expiry date.

Quantum computers will break the encryption banks and governments rely on. Data is being stolen now to be read then, and the EU wants critical systems quantum-safe by 2030. The US CLOUD Act can already reach data US providers hold in Europe.

2 · The data isn't safe · EU post-quantum roadmap · EDPB-EDPS 2019
91%

No plan for the switch.

In a 2025 survey of 1,500 security professionals, 91% had no roadmap for moving to post-quantum cryptography.

2 · The data isn't safe · 2025 survey of 1,500 security professionals
3

Data centers are now targets.

On 1 March 2026, drone strikes damaged three AWS data centers in the UAE and Bahrain, the first publicly confirmed military attack on a hyperscale cloud provider.

2 · The data isn't safe · CNBC, March 2026

Untrusted answers on unprotected data: that is why the institutions with the most to gain from AI still can't use it. That is the market Synapse Net is built for.

It started with an AI you can trust. Protecting it took the rest.

PRISM was the goal. But a trustworthy answer is worthless if the data behind it can be stolen, read or seized, so we built everything around it too: one stack, four parts.

Answers that guess

PRISM understands the question, checks its answer and shows why.

01 · PRISMEncryption quantum computers will break

The Synapse Protocol keeps data sealed, now and after Q-Day.

02 · Synapse ProtocolData that can leave the country

VISUM, inside the protocol, stops it at the border.

03 · VISUMData centers that can be reached

Aegis puts it all inside a mountain, on renewable power.

04 · Aegis
01 · PRISM

Today's AI predicts words. PRISM understands what they mean.

Models like ChatGPT pick the most likely next word, so they can be wrong with total confidence and can't tell you why. PRISM learns the concepts behind a question and how they connect, checks its answer for contradictions before it speaks, and gives you the reasoning with it.

Right answers take less energy: when something doesn't fit, the energy spikes and PRISM corrects itself first.

How PRISM works
Fits the facts: low energyContradiction: energy spikesenergy

Then we had to protect it.

A trustworthy answer is worthless if the data behind it can be stolen, read or seized. So the rest of the stack keeps it sealed, from the wire to the mountain.

See the whole stack
Locked · ML-KEM and ML-DSAIntercepted · still unreadable
Valid visa · passesNo visa · dropped at the edge
1 · Rack2 · Sealed container3 · Security cage4 · Mountain

Seven years of training, down to a day.

The mechanism is proven. What stands between it and a model that talks is compute: on the hardware we have, training would take about seven years. On NVIDIA Blackwell, one run takes about a day.

Seed round open

per training run on NVIDIA Blackwell

Down from about seven years on the hardware we have today, so the architecture can be rebuilt and retested as often as it takes.

The proof works. Compute is what it needs.

seed round, open now. It pays for that compute and the people to use it.

how long training PRISM to basic conversation would take on today's hardware

layers of our own IP

parts, one stack

Companies with one of these layers are worth billions. We are raising $1M.

Every piece of this market is being priced right now. Synapse is the only company putting reasoning you can audit, post-quantum transport and hardware-enforced sovereignty in one stack, and it is raising its first round.

Reasoning beyond LLMsExplainable AIPost-quantum securitySovereign infrastructureValue
AMI LabsWorld models$3.5Bvaluation
GoodfireInterpretability$1.25Bvaluation
SandboxAQQuantum security$5.75Bvaluation
NscaleSovereign GPU cloud$14.6Bvaluation
Synapse NetAll four, one stack$1Mseed round

What each company builds. Synapse's layers are at proof-of-concept and simulation stage. Valuations: AMI Labs pre-money, TechCrunch, Mar 2026 · Goodfire Series B, Feb 2026 · SandboxAQ Series E, Apr 2025 · Nscale Series C, DCD, Mar 2026.

Deadlines

Buyers are on the clock.

Software and AI fall under strict EU product liability from 9 December 2026. High-risk AI must keep logs and explain decisions from December 2027. Critical systems must move to post-quantum cryptography by 2030. Budgets open at deadlines.

Directive (EU) 2024/2853 · EU AI Act, as amended 2026 · EU PQC roadmap
Entry price

This is the proof-of-concept round.

The mechanism already works and is shown live. After production and first pilots, the price of this company will reflect it. The cheapest moment to own it is now.

Where your money goes, and what you own.

Your stake
Equity

In a company that owns three layers of its own IP.

ReasoningCryptographySovereignty
Seed

PRISM to production

The prototype rewritten for NVIDIA Blackwell, with the natural-language decision record.

Seed

Protocol and VISUM

The post-quantum transport in production; VISUM moved from simulation to programmable hardware.

Seed

First pilots

Regulated finance, critical infrastructure and defense, where a wrong answer costs the most.

Next round

Aegis

Our own sovereign node in Norway.

See the investor pageGet the investor deckGet the investor deckGet the investor deck

Nothing on this website is an offer to sell or a solicitation to buy securities. Investing in an early-stage company is risky, including the risk of losing the entire investment.

Built for institutions that get audited

Banks, grid operators and defense ministries must explain every decision and protect every byte. Every part of the stack was designed for that room.

Sealed

Post-quantum from day one

ML-KEM and ML-DSA, the algorithms NIST standardized in August 2024.

Data stays in its country

VISUM drops any packet without a valid visa at the edge of the network.

A record of every decision

PRISM writes down how each answer was reached, in plain language.

Out of physical reach

Aegis is planned 700 meters inside a Norwegian mountain, on NATO territory.

In the model makers' own words

We could tell you chatbots don't reason. The labs that build them have already published it. These are the lines we underlined.

OpenAI researchers
Kalai, Nachum, Vempala, Zhang
arXiv 2509.04664
4 Sep 2025

“Language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty.”

Why Language Models HallucinateRead the source
Anthropic
Interpretability research
Tracing thoughts
27 Mar 2025

“Claude, on occasion, will give a plausible-sounding argument designed to agree with the user rather than to follow logical steps.”

Tracing the thoughts of a large language modelRead the source
Anthropic
Alignment science
Faithfulness
3 Apr 2025

“A substantial majority of answers, then, were unfaithful.”

Reasoning models don't always say what they thinkRead the source
Apple
Machine learning research
Illusion of thinking
Jun 2025

“Frontier LRMs face a complete accuracy collapse beyond certain complexities.”

The Illusion of ThinkingRead the source
Llion Jones
Co-author of Attention Is All You Need
TED AI
Oct 2025

“I'm explicitly now exploring and looking for the next big thing.”

Sakana AI's CTO says he's 'absolutely sick' of transformers (VentureBeat)Read the source

It started with one engineer who saw the flaw in how chatbots work.

GM

Giuliano Maroulis

CEO & Lead Architect

Architect of the Synapse stack. When GPT-3 launched he saw the problem as architectural and spent three years designing the stack from first principles. He built the PRISM prototype himself.

LF

Lorenzo Festa

Chief Operating Officer

Operator and founder. He took a client from zero to over €1M in annual revenue. At Synapse he leads strategy, funding and go-to-market across Europe.

Read our story

Learn

Short, direct answers to the questions behind sovereign AI.

View all articlesView all articlesView all articles
PRISMLearn

Can LLMs reason? No. They predict what reasoning looks like.

A large language model predicts the next word from probabilities. Why that looks like reasoning, why it is not, and what the model makers' own research shows.

Read article
PRISM

Who invented the transformer, and why Google did not win with it.

PRISM

Why does AI hallucinate? Because nothing inside it checks.

Synapse Protocol

Harvest now, decrypt later: the data being stolen for Q-Day.

VISUM

The US CLOUD Act and European data: the conflict no contract fixes.

What people ask us first.

Straight answers to what investors, partners and engineers ask us first.

No. ChatGPT and other large language models write by predicting the next word, and nothing inside them checks whether the answer is true. PRISM is an energy-based model: it learns concepts and how they relate, scores a whole answer before it speaks, and corrects itself when a contradiction makes the energy spike.

Yes. In the live demo you can steer PRISM, watch it contradict illogical input, and see the full spectrum of what happened while it reasoned and how it classified each step.

A model that gives every possible answer a score called energy. Low energy means the answer fits; high energy means something contradicts. Reasoning becomes finding the answer with the lowest energy. More on energy-based models.

Yes. A language model can still do the talking while PRISM does the checking.

AI whose model, data, transport and infrastructure stay under the control and law of one jurisdiction. Read our definition.

A stack for institutions that need AI they can trust and control: PRISM for reasoning with an audit trail, the Synapse Protocol for post-quantum transport, VISUM for jurisdiction-locked networking, and Aegis for sovereign compute.

Aegis is planned inside Lefdal Mine Datacenter, a former mine in Måløy, Norway, on NATO territory. Agreements are in progress.

No. If the operator is subject to a foreign law such as the US CLOUD Act, the data can be compelled regardless of where the server stands.

A $1M seed round to take PRISM, the Synapse Protocol and VISUM from proof of concept to production.

Mostly compute and the people to use it. On the hardware we have, training PRISM to basic conversation would take about seven years; on NVIDIA Blackwell, one training run takes about a day.

The live demo, the system blueprints under NDA, and the go-to-market plan. Write to investors@synapsenet.io.

No. Nothing on this website is an offer to sell or a solicitation to buy securities. Investing in an early-stage company is risky, including the risk of losing the entire investment.

Giuliano Maroulis, CEO and lead architect, who designed the stack and built the PRISM prototype, and Lorenzo Festa, COO, who leads strategy, funding and go-to-market. Our story.

No. Synapse Network Solutions is not affiliated with Synapse Financial Technologies Inc.

Still have questions? Write to the founders

The proof-of-concept round closes once. Be in it.

Live demo, technical brief and data room for qualified investors.

A live demo of the proof of concept
Technical brief and data room under NDA
$1M seed round, open now
Get the investor deckBook a live demoBook a live demoBook a live demo

Nothing on this website is an offer to sell or a solicitation to buy securities.