Ramp's internal AI workspace, Glass, burned through $1.5 million in tokens in a single week in May 2026. The figure arrived tucked inside a product announcement — a footnote to a launch, not a headline. But it is the most consequential number in the entire disclosure, and it deserves more attention than the product itself. A weekly inference bill of that size is larger than the annual compute budget of many mid-stage startups. It tells us where the true cost of "AI-native" enterprise software is migrating: away from the model layer, and into the retrieval and orchestration machinery that sits between a company's accumulated knowledge and its daily output.
I have spent the better part of a decade watching the same pattern repeat across different technologies. In 2020, I spent three weeks modeling impermanent loss on a USDT/ETH pair, documenting how algorithmic stablecoins quietly redistributed wealth from retail to whales. The mechanism was invisible to most users — it lived in the spread, in the fee curve, in the timing of liquidity. The same invisibility now surrounds enterprise AI. Everyone admires the output; almost no one audits the cost structure that produces it.
Glass is not a model breakthrough. It is a knowledge-engineering product, and the distinction matters more than Ramp would like to admit.
The product is straightforward enough. Ramp, a fintech company best known for corporate cards and expense management, built an internal workspace that stitches together the tools its employees already use — Slack, Notion, Linear, Google, Gong — and layers a conversational interface on top. A small internal team of what the company calls its "most skilled AI users" had their workflows — prompts, tool chains, knowledge bases — distilled into a shared environment that any of Ramp's 323 employees can now use. The company reports 6,500-plus edits, 20,000 records entered, and, most notably, a 30 percent reduction in its own token bill after migrating its work into Glass.
That last number is the one to sit with. A 30 percent reduction in inference cost does not come from a better model. It comes from caching, from prompt discipline, from routing simple tasks to cheaper models and reserving the expensive ones for genuine complexity. These are the unglamorous mechanics of applied AI, and they are where the real engineering lives. The model is a commodity; the orchestration is the moat.
Let us do the arithmetic the announcement avoided. At prevailing frontier-model pricing — roughly $10 to $30 per million tokens across a mix of input and output — $1.5 million a week implies somewhere between 50 and 150 million tokens processed weekly across a 323-person company. That is hundreds of thousands of tokens per employee, every week. For a retrieval workload, that volume is not a sign of productivity; it is a sign of inefficiency waiting to be engineered away. The 30 percent reduction is not the end of that curve. It is the first step on it. At scale, that is a margin question, not a feature question.
For readers who track on-chain infrastructure, this should feel familiar. The value in a decentralized exchange was never the smart contract template — those are copyable in an afternoon. The value was in the liquidity, the routing, the accumulated depth that made the venue worth using. Glass applies the same logic to organizational knowledge. Ramp's prompts are copyable. Its knowledge graph — the tangled, permissioned web of who knows what, and which document answers which recurring question — is not.
There is a quieter parallel, too. The entire premise of Glass — start from your company's existing knowledge, not from a blank slate — is a retrieval problem dressed as a generation problem. Retrieval-augmented generation is not new; it is the backbone of every serious enterprise deployment. What Ramp has done is industrialize it, and in doing so, surface the cost that retrieval hides. Every query against a knowledge base is a query against an index. Every index needs maintenance. The $1.5 million week is the bill for that maintenance, rendered in tokens.
Coming from cross-border payments, I recognize the pattern. In 2024 I analyzed 12,000 remittance transactions and found that stablecoins cut settlement from five days to fifteen minutes while reducing costs by 40 percent. The savings did not come from a better dollar. They came from removing intermediaries and collapsing the number of hops between sender and recipient. Glass is doing the same thing to organizational knowledge — collapsing the hops between a question and its answer. The efficiency is real. So is the question of who controls the rails. Between the wire and the wallet, there is a void — and Glass is quietly trying to fill a version of it inside the enterprise.
This is where my current work — auditing decentralized compute networks here in Lagos — becomes relevant. The central promise of those networks is cost: aggregate idle GPU capacity, route inference through a permissionless market, and undercut the hyperscalers. For a company burning $1.5 million a week on tokens, that pitch is not academic. It is existential. Yet the same audit discipline I applied to ERC-20 contracts in 2017 applies here. Decentralized compute solves the pricing problem and inherits the verification problem. Who guarantees the model actually ran the way it claimed? Who audits the output when the compute node is anonymous? Glass sidesteps these questions by staying centralized — and that is precisely why it can charge for trust.
The centralized workspace wins on trust and loses on cost. The decentralized network wins on cost and loses on trust. Neither has resolved the trade-off, and anyone claiming otherwise is selling something.
Here is the contrarian reading. I see the pattern before it becomes a trend, and this one carries the unmistakable fingerprint of a narrative in search of a market. Ramp frames Glass as evidence of an "AI-native company" — a culture, not a tool stack. That framing is seductive, and it is also the same narrative architecture I have watched hollow out the interoperability space. The "omnichain app" promised that users would care how many chains your contracts touched. They did not. The "AI-native company" promises that organizational knowledge will flow freely across every team. It will not — not without confronting the permission problem that Ramp's announcement conspicuously omits.
Integrating Slack, Notion, and Gong means integrating their access controls. A knowledge base that answers every question is also a knowledge base that leaks every answer. The announcement says nothing about encryption, nothing about audit logs, nothing about how a junior employee's query is prevented from surfacing a board-level document. These are not footnotes. They are the load-bearing walls of the entire product, and they are invisible.
The technical difficulty here is genuinely underrated. Permission-aware retrieval is one of the hardest problems in applied AI, because it forces the system to reason about identity before it reasons about content. A blockchain solves the analogous problem with cryptographic proofs — every actor is a keypair, every access a signature. Glass solves it with something far more fragile: organizational trust, expressed through integrations that were never designed to interoperate. The two approaches are not competitors. They are mirror images, and the gap between them is exactly where the next generation of infrastructure will be built.
DeFi promised freedom; it delivered a mirror. Enterprise AI promises collective intelligence; it may deliver the same silos, only faster and better-indexed.
The 30 percent cost reduction is real, and so is the automation case — Ramp reports Salesforce feature requests growing from 300 manual entries to 20,000 captured automatically, with 5,000-plus emails dispatched. But it is Ramp's own data, from Ramp's own employees, presented by Ramp's own co-founder. I have audited enough distribution logic to know that internal benchmarks and external reality are separated by a wide, unflattering gap. Survivorship bias is not a flaw in the announcement; it is the announcement.
What matters now is not whether Glass succeeds as a product. It is whether the organizational knowledge layer becomes the next contested territory in the AI stack — and whether decentralized alternatives can compete on trust before the incumbents close the cost gap. The token bill tells us the market is real. The silence on permissions tells us the hard part has not begun.
We map the flows, but the ocean remains unmapped.
