A new VSCode extension started circulating this week that quietly reroutes AI coding requests through NEAR AI Cloud instead of a hyperscaler's backend. The pitch is narrow and sharp: your proprietary code stays private even while a model reads it. That is the headline. But sitting in my terminal past midnight, watching the extension handshake, one question kept surfacing — who actually pays for this, and what does NEAR's token get out of it? Chasing the alpha, one block at a time, I went digging, and the story is smaller and stranger than the press release suggests.
NEAR Protocol is a sharded proof-of-stake L1 that has been running since around 2020. It is not a scrappy upstart. The chain has survived multiple cycles, and its credibility in the AI conversation rests on a genuinely rare credential: co-founder Illia Polosukhin is one of the authors of "Attention Is All You Need," the 2017 paper that gave the world the transformer architecture. That single fact gives NEAR something most AI-branded crypto projects can never buy — narrative legitimacy. When NEAR says it is building AI infrastructure, it is not borrowing the story. It helped write the original chapter.
The product in question, NEAR AI Cloud, sits at an unusual layer. It is not a chain upgrade. It is a service that pairs NEAR's L1 with a private inference layer and then wraps that in a developer distribution channel — a VSCode extension that plugs into GitHub Copilot. The promise to developers is twofold: privacy, meaning your code is not exposed to a third-party model provider, and flexibility, meaning you choose how and where inference runs. On paper, that combination targets the exact anxiety enterprise developers feel every time they paste a proprietary function into an AI assistant.
Here is where I have to slow down and mark the boundaries. The original announcement gave us four facts and almost nothing else. No technical whitepaper. No latency benchmarks. No throughput numbers. No cost per token. No disclosure of how the "private" part is actually implemented. From the front lines of the hype cycle, I have learned that the missing numbers are usually the interesting ones. And we are reading this in a sideways market, where narrative is cheap and verifiable usage is scarce — which makes the gaps even louder.
Let me start with the technical reality, because this is where the marketing gets slippery.
This is a distribution-layer innovation, not a technical paradigm shift. NEAR is not inventing a new inference algorithm. It is taking an existing capability — private inference — and slotting it into the single most valuable piece of real estate in software right now: the developer's editor. That is a smart move, but it is integration, not invention. The competitive moat is the entry point, not the math.
The word "private" is doing enormous unexamined work here. There are at least three ways to build private inference, and they carry wildly different trust assumptions. The first is a trusted execution environment, or TEE — hardware-level isolation like Intel SGX or the confidential computing modes on modern Nvidia chips. It is fast, but it means you are trusting the chip vendor, not cryptography. The second is a purely cryptographic route — fully homomorphic encryption or multi-party computation — which is trust-minimized but carries brutal performance penalties. The third is federated learning, which keeps data local but changes what you can actually compute. NEAR did not say which one it uses. That silence matters, because "private" in a TEE world means "private as long as you trust Intel," and that is a very different product from "private because the math says so."
I have audited enough of these claims to know the tell. When a team leads with the privacy outcome and omits the privacy mechanism, it is usually because the mechanism has a tradeoff they would rather not foreground. I am not accusing anyone of bad faith. I am noting that the burden of proof is on the product page, and the product page is empty.
Then there is the integration-depth question. "Brings private inference to GitHub Copilot" could mean two very different things. Option one: NEAR replaces Copilot's default inference backend entirely, a deep, high-leverage integration. Option two: NEAR ships as an optional plugin or model source that developers must deliberately select. The wording leans toward option two. That distinction decides the entire addressable market. A default backend reaches every Copilot user. An optional plugin reaches the tiny fraction of developers curious enough to install it. One is a distribution win. The other is a hobby project with a press release.
There is also a performance black hole. Enterprise buyers do not adopt inference infrastructure on privacy alone; they adopt it on privacy plus acceptable latency plus predictable cost. Without published benchmarks, no serious engineering team can run a bake-off against Azure or Bedrock. That absence does not just weaken the sales pitch. It suggests the numbers may not yet be flattering.
Now the token. This is the part that keeps me up.
The announcement claims NEAR token holders benefit from a staking model integrated with the AI service. That is a single sentence carrying the weight of an entire thesis, and it tells us nothing about the mechanics. In crypto, "staking model" is a phrase that can mean three completely different things. It can mean the token is a payment medium, where enterprises pay for inference in NEAR, creating genuine demand. It can mean the token is staking collateral, where you lock NEAR to earn a share of service revenue, which only works if the revenue is real. Or it can mean the token is pure governance theater, a narrative decoration with no economic binding at all.
The single most important unanswered question in this entire story is whether AI service revenue actually flows back to stakers, or whether the "yield" is just new inflation dressed up as AI income. If the rewards come from fresh token issuance, this is not an AI business. It is a subsidy program wearing an AI costume. I have watched this exact pattern play out across dozens of "utility staking" launches, and the tell is always the same: real revenue is disclosed loudly and early, while subsidy-based rewards are described only in vague, aspirational language. NEAR's description is vague and aspirational.
Zoom out to the competitive field and the pressure sharpens. Bittensor runs a decentralized network of inference subnets with trust-minimized verification. Akash and io.net compete on raw GPU markets. Render is expanding from rendering into compute. And behind all of them sit the hyperscalers — Azure, AWS Bedrock — with mature tooling, massive scale, and enterprise sales teams that already have the chief information officer's phone number. NEAR's differentiation is a composite: L1 plus developer tooling plus founder credibility. That is a real combination. But a combination is not a network effect, and network effects are what actually win infrastructure markets. Speed is the only currency that matters, and here the speed is on the side of the incumbents.
On the ecosystem side, the distribution signal is genuinely the most valuable part of this announcement. Plugging into Copilot and VSCode is zero-friction distribution. It puts NEAR's service one keystroke away from tens of millions of developers without a single ad or conference booth. But that reach comes with a structural catch: it is entirely dependent on Microsoft's platform. If GitHub changes its policy on third-party inference sources — and it can, unilaterally, at any time — NEAR's distribution channel evaporates overnight. Building your growth engine inside someone else's walled garden is a strategic bet, and it is not a safe one.
There is a compliance layer too, and it cuts both ways. Privacy is the selling point, but if NEAR courts European enterprise customers, the same privacy promise collides with GDPR, data-residency rules, and potentially cross-border encryption export controls. "Private inference" can be a compliance asset or a regulatory question mark, depending on which jurisdiction is asking.
Here is the angle almost nobody is reporting. Everyone is reading this as an AI story. It is not. It is a talent-and-narrative story about a chain trying to escape a structural trap that every mature L1 eventually faces: staking yields that come from inflation rather than real demand. When your block reward is the product, you are running a treadmill, not a business. NEAR's founder credibility is a scarce asset, and the AI pivot is the most credible way to convert that asset into new, non-inflationary demand for the token. That is the real bet.
The contrarian read is that this announcement is a positioning move, not a product milestone — and that is not necessarily bad. It strengthens the "NEAR equals AI chain" mental slot in a way that is very hard for competitors to dislodge, precisely because of the transformer paper lineage. You cannot fake that origin story. But positioning without verifiable usage eventually curdles into disappointment. If NEAR cannot show measurable inference volume or paid usage within six to twelve months, the narrative premium unwinds, and the same founder credibility that lifted it becomes the yardstick it gets measured against. Turning red candles into green lessons means understanding that a strong story buys you time, never results.
So watch the numbers, not the press release. The moment NEAR publishes inference call volume, paying customers, or a clear revenue-to-staker flow, the thesis gets real. Until then, this is a directional signal — a chain planting for spring while the market chops sideways. Surviving the winter to plant for spring is noble. But spring only arrives for the ones who can prove the seeds actually sprouted. The sprint never stops, only the pace — and right now, NEAR is running on reputation. The clock is already ticking.