Behind DGrid: From Stony Brook Research to Verifiable AI

The number of usable AI models has grown exponentially, from major firms such as OpenAI, Anthropic, and Google AI to various smaller, open-source projects. Using any of them, though, still means a separate sign-up with each provider, their pricing, and trusting that their servers and results are genuine, which essentially raises a “black box” issue, wherein there is no transparency into the authenticity of the inference being run.

DGrid is a project aimed at building the infrastructure to run AI models, enabling an economy that gives users better access to the latest models and allows them to contribute compute power and intelligence in return. Much of this also draws on a recent Twitter Spaces appearance by founder Alex Ding, who spoke on the product, the team behind it, and what's coming next.

DGrid’s core product, the AI Gateway, gives developers a single API endpoint to call 200+ LLMs from various providers, with access and payments handled via smart contracts rather than a single company's billing system. "Inference" simply means running a trained AI model on a new input to produce an output, which is essentially what a chatbot does each time it answers a prompt.

DGrid states that its purpose is to solve three problems in the AI market:

1. Interface Fragmentation - each model provider runs its own API and pricing scheme

2. Value capture by Centralized Platforms - as per DGrid’s statements, this squeezes out model creators and prompt engineers

3. Lack of verifiability in AI outputs - this is documented as a “black box” issue.

Its proposed solution combines three components: a routing and verification layer, an open model marketplace, and an AI-DAO, which is essentially just a token-based governance system.

Who is behind DGrid?

Alex Ding - Co-founder and CEO

Alex completed graduate studies in Technology and Society Management and Computer Science at Stony Brook, with 10+ years in blockchain operations and five in machine learning. He is also the corresponding author across DGrid's three published PoQ papers: the original cost-aware scoring design, a follow-up proposing a hybrid TEE-based verification architecture, and a later paper on hardening PoQ against adversarial evaluators- a sustained research thread behind the product, not a one-off credential.[4] On the Spaces call, Ding put the wider team at 12 full-time staff, mostly Stony Brook alumni.

Aaron Chan - CTO

Aaron holds a B.S./M.S. in Mathematics from CUHK, a Ph.D. in Applied Mathematics and Statistics from Stony Brook, five-plus years in AI, and eight-plus in blockchain research. He is a co-author on the same three papers and, given his doctoral focus, is the more direct technical fit for that authorship, which is consistent with the "Ph.D.-level team" claim used in earlier press coverage, though that describes one confirmed member, not the whole team. Ding's own Stony Brook credential is graduate study, not a doctorate, and Khaldi's bio lists none.

Chaymaa Khaldi - CMO

Chaymaa is credited with more than three years of global marketing and brand management experience across technology companies. According to DGrid, Chaymaa served in senior marketing roles at multiple prominent tech firms and now focuses on crafting and driving innovative go-to-market initiatives that boost DGrid’s brand equity and expand its market presence.

Routing and Execution of Requests

Requests to the AI Gateway are compatible with OpenAI’s software development kit (SDK), so a developer can point existing code at DGrid’s endpoint with minimal changes. Behind that endpoint sits a network of independently operated DGrid Nodes. These node operators supply GPU hardware and choose which models to host, ranging from lightweight models with 7 billion parameters to those with more than 70 billion parameters, depending on their hardware.

Nodes report their own latency and Compute Unit (CU) usage, which is a standardized measure of the inference complexity tied to model size and token count, back to the network for routing and pricing purposes. Routing and per-task fees consequently rest on performance data that the nodes report about themselves. It is worth noting that the official DGrid documentation does not provide a way to verify the accuracy of self-reported data. That is a separate question from Proof of Quality, which only verifies the AI’s output, not whether a node is honest about its own performance metrics.

How does Proof of Quality (PoQ) work?

Because inference occurs on nodes that DGrid does not control, the network needs to verify that a node ran the correct model and met a quality bar. PoQ scores outputs on accuracy against reference answers, consistency across nodes, and format compliance. Scores are written on-chain, letting a user verify provenance without re-running the task. On the Spaces call, Alex filled in how this runs day-to-day: an automated check samples one in every 100 outputs, and a Singapore-based human review team separately spot-checks results against the model provider's own API, a mechanism he called still maturing, given how varied real queries are next to the training data.

PoQ itself is a statistical consensus method, not a cryptographic proof; correctness comes from evaluator agreement, not mathematical verification. A 2025 paper co-authored by DGrid's own research team acknowledges this, noting Standard PoQ trades cryptographic certainty for subjective scoring, leaving the network exposed to nodes that quietly swap in a cheaper model. That paper proposes a hybrid design that combines trusted hardware with cryptographic spot checks. DGrid's docs mention a "hybrid verification architecture" but don't explain what makes it hybrid, so it's unclear whether the label refers to the paper's design or something else entirely.

Rollout Status

Validator staking, which DGrid says gives PoQ real economic weight, went live August 25, 2026. (DGrid's roadmap places full on-chain governance, an Agent launchpad, the Model & Agent Market, DGrid Scan, and AI DAO 1.0 all in Q3-Q4 2026, the period the network is in now, with no confirmation that any have been activated. Separately, DGrid's docs label its three PoQ papers "peer-reviewed," though all three are arXiv preprints by DGrid's own team, and arXiv doesn't peer review submissions. On the Spaces call, Alex singled out the marketplace specifically as the piece to watch over the next few months, the next major launch still ahead, by his account.

Conclusion

Put together, DGrid is essentially a marketplace with two sides. On one side are users with spare GPU hardware, and people who have built an AI model or agent worth offering; both are listed on the network and earn $DGAI for their contributions. On the other side are developers and users who want AI output without opening a separate account with every provider that makes it (i.e. OpenAI, Anthropic, etc.). $DGAI is the network’s native currency that moves between the two. It is what buyers spend, and what suppliers earn.

The case for why this should matter, as DGrid frames it, is that AI access today runs through a small number of companies that set the price, control who gets in, and give no way to check that a result actually came from the model paid for (i.e. the black box issue). DGrid aims to let many independent suppliers compete for that business instead of one company, with a verification layer meant to give buyers something to verify their result against

Whether that case holds up rests less on the API or tokenomics, and more on whether the verification layer and its governance end up doing what a single trusted provider currently cannot, which will only be proved once the network is stress-tested in real-time markets.