Inheritance is a feature until it becomes a trap.
A $400 million check. No technical whitepaper. No team announcement. No public repository. Just a press release and a vision: “the free World Wide Web for AI.” That is the launch of Current AI – a non-profit open infrastructure initiative backed by Google and the French government.
Let me be blunt. As someone who has audited smart contract upgrades for the Ethereum Classic hard fork, standardized interest rate interfaces during DeFi Summer, and dissected the royalty reentrancy bug in a leading NFT marketplace, I know what an incomplete launch looks like. This one screams “announcement without execution.” But it also signals something deeper: the AI industry has entered the infrastructure layer of geopolitical struggle.
Context: The Empty Vessel
The announcement contains almost no technical information. Current AI is described as a “non-profit organization” aiming to build an open, decentralized, non-profit AI resource layer – a “free World Wide Web for AI.” The supporters are Google (through both corporate backing and likely cloud credits) and the French government (through policy support and potential public funding). The initial capital commitment is $400 million.
That number seems large until you realize it is a fraction of the cost required to train a frontier model. GPT-4’s training cost was estimated at $100-200 million. A single training run for a 1-trillion-parameter model on 10,000 GPUs can exceed $50 million in electricity and hardware depreciation. $400 million is not a research budget. It is a strategic down payment.
The real asset is not money. It is legitimacy. Google wants a non-profit shield to weaken the Microsoft/OpenAI alliance. France wants digital sovereignty. The open source community wants a neutral platform. Three different motivations converging on a single narrative: “free AI for everyone.”
Core: Dissecting the Open Infrastructure Promise
Let me deconstruct what a “free World Wide Web for AI” actually requires. I will examine it through the lens of technical architecture, not marketing.
First, the infrastructure layer must include five components: 1. Compute orchestration: A universal scheduler that can aggregate GPU clusters from Google Cloud, French supercomputers, and donor PCs. 2. Data layer: Standardized datasets with provenance tracking, licensing, and quality scoring. 3. Model registry: A repository of open models with versioning, compatibility checks, and execution environments. 4. Interoperability protocol: A common API format that allows any model to be called from any application, regardless of training framework. 5. Identity and governance: A decentralized system for attribution, access control, and dispute resolution.
The engineering challenge is massive. Building a system that can schedule a training job across Google Cloud TPUs, French HPC clusters using AMD MI300X, and spare Nvidia H100s from a university requires a distributed operating system that does not exist today. The network latency alone can destroy training efficiency. My experience auditing the Compound protocol taught me that standardizing interest rate models across different lending pools required months of consensus-building and multiple ERC extension proposals. That was for a single financial primitive. Current AI intends to standardize the entire AI stack.
Second, the “free” label is deceptive. Access to compute will never be free at scale. The electricity and hardware costs are real. The model is more likely a market maker: a low-friction platform that charges minimal overhead but relies on surplus compute donations and government subsidies. This is exactly how the World Wide Web evolved – free at the point of use, but paid for by ISPs, universities, and advertising. The problem is that AI compute is five orders of magnitude more expensive than web hosting.
Third, the security surface is enormous. An open AI infrastructure allows anyone to deploy models. Malicious actors can upload models that generate deepfakes, exploit vulnerabilities, or exfiltrate user data. During the OpenSea audit, I discovered a reentrancy bug in the royalty module because the contract assumed off-chain royalty enforcement was safe. Current AI will face the same assumption: that open equals safe. It is not. The platform must implement on-chain verification for every model execution – a feat that requires a trustless execution environment akin to a smart contract runtime for AI.
Fourth, the governance paradox. Who decides which models are removed? Who enforces regulatory compliance under the EU AI Act? The French government, as a backer, will demand censorship of hate speech and extremist content. Google will want to protect its brand. The non-profit board may become a battleground. I have seen this pattern before: the ETC hard fork audit revealed that even technical fixes can be politicized when governance is unclear.
Contrarian: The $400 Million Blind Spot
Everyone is celebrating the open infrastructure. I see a trap.
Trap 1: The governance vacuum. No one has defined how Current AI will make decisions. If Google holds the cloud credits, it holds bargaining power. If the French government provides the data-centers, it holds veto power. The “non-profit” label is a legal fiction that can be captured by the largest contributor. The same dynamic played out with the Linux Foundation – initially independent, but eventually dominated by corporate members. The difference is that Linux kernel was a finished product. Current AI is a blank slate that needs continuous funding. Any contributor can threaten to withhold resources to steer the direction.
Trap 2: The compute landgrab. $400 million is not enough to build a competitive compute network. Meta spent over $20 billion just on GPU purchases in 2024. Current AI’s capital is a rounding error. The only way to get compute at scale is to rely on donations and government subsidies. But who donates compute? Companies that want a seat at the board. This creates a perverse incentive: donate compute, receive influence, then lobby for protocols that benefit your own commercial interests. The “neutral” infrastructure becomes a puppet for the largest donors.
Trap 3: The fragmentation risk. The open web succeeded because of a few key protocols: HTTP, HTML, TCP/IP. There was no competition. In AI, we already have multiple competing standards: OpenAI’s API, Google’s Vertex AI, HuggingFace’s Hub, and on-chain decentralized compute networks like Bittensor. Current AI will add another layer. Unless it can achieve critical mass by attracting existing players to migrate, it will become just another silo. My work on institutional custody standards for AI-crypto hybrids taught me that interoperability requires painful coordinated upgrades. In a non-profit with three primary funders, coordination is political, not technical.
Trap 4: The security debt. An open platform that allows anyone to run arbitrary models is a honeypot for attacks. Models can be poisoned during training to produce back doors. Inference can be used to steal private data. Without a smart contract-style execution sandbox and immutable audit trail, Current AI will become a vector for large-scale harm. The Terra-Luna collapse taught me that feedback loops that look stable on paper can explode when economic incentives align with abuse. The same applies to open AI: the incentives for bad actors are high, and the infrastructure’s non-profit nature makes it slow to respond.
Trap 5: The talent exodus. Building a global AI infrastructure requires top-tier engineers. Those engineers can earn $1 million+ at Google or OpenAI. A non-profit cannot compete on salary. It will rely on mission-driven talent. But after the dot-com bubble burst, many non-profit open source projects struggled because the best contributors left for high-paying jobs. Current AI will face the same challenge: the people who can build this system are the same people who can earn ten times more building proprietary systems.
Takeaway: A fork in the road
Current AI is not a technology project. It is a political, economic, and governance experiment. If it succeeds, it will be because the founders designed a governance system that prevents capture by any single stakeholder, distributes decision-making to a broad community, and sets clear security boundaries. If it fails, it will be because the $400 million ran out before the system gained enough adoption to become self-sustaining, or because the governance collapsed under the weight of competing interests.
I have audited projects that looked good on the surface. The Compound standard I helped draft took years to penetrate the ecosystem. The Terra-Luna mechanism I analyzed afterward was mathematically elegant but socially fragile. Current AI is the same: a beautiful concept that will live or die by its execution.
The question is not whether open AI infrastructure is desirable. It is whether we can build it without repeating the mistakes of the past. Inheritance is a feature until it becomes a trap. Execution is final; intention is merely metadata.
