Last month a GPU landlord started giving away software. CoreWeave's new Forge platform โ a full agent development stack spanning sandbox execution, observability, model registry, distillation, and a research agent called ARIA โ lists at $60 per month for teams up to fifty seats. Run the arithmetic. That is $1.20 per head. LangSmith charges roughly $39 per seat. Braintrust sits in the same neighborhood. W&B, before CoreWeave bought it for $1.7 billion, never charged $1.20 for anything.
I have spent the last eight years treating vendor pricing pages the way I treated gas-price spikes during the CryptoKitties congestion in December 2017 โ refresh, screenshot, timestamp, cross-reference against block data. The $60 number is not a pricing decision. It is a confession. CoreWeave is telling the entire market that the control plane for AI agents has zero pricing power, and that it intends to run software at a loss to defend something else: GPU utilization.
That confession matters far beyond the AI infrastructure crowd. It is the clearest signal yet that the decentralized physical infrastructure (DePIN) compute thesis โ Akash, Render, io.net, Aethir, Gensyn โ is fighting on the wrong axis. And it is a signal that every crypto founder who has ever pitched "decentralized GPU" to a token holder needs to read carefully.
Context: How a GPU Renter Became a Platform
CoreWeave's trajectory is well-documented and verifiable through public filings, which is where I always start. The company IPO'd in March 2025. It reported roughly $1.9 billion in 2024 revenue, guided toward the $5 billion range for 2025, and carried a backlog in the vicinity of $30 billion. Those are SEC numbers, not press-release numbers. I trust them the way I trust a finalized block โ with the understanding that filings can be restated, but not invented.
The strategic moves that produced Forge are traceable. In March 2025, CoreWeave announced the acquisition of Weights & Biases for approximately $1.7 billion, closing in May. That deal brought the industry's most widely adopted experiment-tracking and observability layer under a compute provider's roof. Later in 2025, CoreWeave acquired OpenPipe, a company known for distilling large models into cheaper small ones. And the entire stack sits on top of a deep, unusually tight relationship with NVIDIA โ investment, priority supply, and a capacity backstop agreement reported in the $6.3 billion range, under which NVIDIA effectively commits to purchase unsold capacity.
Read that again. A company whose core asset is scarce GPU supply has signed an agreement with its own supplier to guarantee demand. That is not a normal vendor relationship. That is a strategic dependency dressed as a moat.
Forge is the logical next step. It packages W&B's tracking, OpenPipe's distillation, and a sandbox execution layer into a single lifecycle: Run, Observe, Curate, Improve, Evaluate. Five stages. It is a rename of the standard MLOps loop โ the same loop MLflow, LangSmith, and every serious observability vendor has shipped for years.
Before I go further, I owe you the same disclosure I give on every on-chain investigation. I cannot independently verify the performance claims in CoreWeave's Forge announcements. The "20% improvement in fault detection," the "10x lower cost," the "3x faster sandbox startup with Vera CPU" โ these are vendorๅฃๅพ, single-source numbers. My rule with single-source numbers is the same as my rule with unconfirmed transaction hashes: discount by 50% to 70% until a second chain confirms it. So read everything below with that haircut applied.
Core: The Five-Stage Lifecycle Is a Rename, Not a Reinvention
Let me be precise about what Forge is and is not. It is not a model architecture. There is no attention-mechanism innovation, no novel training objective, no new loss function. The technical content lives entirely in two places: the sandbox isolation mechanism and the CPU-level startup optimization. Everything else is repackaging.
Agent Lens maps directly onto W&B Weave and LangSmith's agent tracing. Registry maps onto MLflow Model Registry and Hugging Face Hub's lineage versioning. Model Distillation is a service-ized teacher-student pipeline. Sandboxes are an isolated execution environment, the same category as E2B, Modal, and Daytona. Not one of these is an industry first. The innovation tier is combination-level plus engineering-level. That is not an insult. Most enterprise value in history has been created by combination, not invention. But it is a fact that the marketing does not want you to internalize.
The Sandbox Question: Kata vs. Firecracker
Here is where it gets technically interesting, and where I lean on the closest thing I have to a personal benchmark. In 2020, during DeFi Summer, I deployed small capital across Uniswap and Compound specifically to feel impermanent loss and slippage in my own wallet rather than read about it. I later applied the same trial-based method to sandbox environments โ spinning up Firecracker microVMs and container-based runners to measure cold-start latency with my own stopwatch, because published benchmarks are marketing.
CoreWeave chose Kata Containers for its sandboxes. Kata provides VM-level isolation, OCI-compatible, with each sandbox running its own kernel. DigitalOcean, for its Agent Droplets, chose Firecracker microVMs. This is a genuine architectural fork, not a cosmetic one.
Kata's advantage is isolation strength and compatibility. Its disadvantage is startup overhead โ typically heavier than Firecracker, whose cold-start can reach the ~125 millisecond range. Firecracker was built by AWS to be brutally fast and thin. Kata was built to be a secure boundary. If you are selling to regulated enterprises and healthcare, you pick Kata. If you are selling to cost-sensitive developers, you pick Firecracker.
So why does CoreWeave claim Vera CPU delivers 3x faster sandbox startup? Because the architecture choice created a latency debt, and hardware is the only lever left to pay it down. This is the tell. The 3x figure is almost certainly a selective benchmark โ likely measured on memory snapshot restore, per-core efficiency, or against an older x86 baseline. VM startup bottlenecks usually sit in image pull, memory snapshot restore, and kernel initialization. A single CPU architecture swap does not produce 3x end-to-end improvement across a general workload. I have benchmarked enough runners to be confident about this. Confidence: low on the claim, high on the skepticism.
Distillation Is Data Gravity, and Nobody Is Talking About It
Here is the part of the Forge story that I have not seen a single mainstream outlet frame correctly. Model Distillation plus Agent Lens is not a productivity feature. It is a data-capture device.
Think about the mechanics. Agent Lens records production traces โ every prompt, every tool call, every failure, every successful reasoning chain. Model Distillation trains smaller, cheaper specialist models. Put them together and what you have is a pipeline that ingests a customer's proprietary production behavior and converts it into a distilled asset. The question that Forge's terms of service will eventually have to answer: are customer traces being used to improve CoreWeave's own models, and can a customer opt out?
This is the same data-sovereignty question that DePIN projects claim to solve by design. On a decentralized network, the argument goes, your data never leaves your control because the compute is verifiable and the storage is content-addressed. On Forge, your data becomes the training set for a model you may not own. The crypto-native reader should find that trade uncomfortable. Whether they will notice it before signing the enterprise contract is another matter.
There is a second tension buried here. OpenPipe's entire commercial premise is replacing large models with small ones โ that is a cost-reduction story that, on its face, reduces GPU consumption. CoreWeave sells GPU time. So how does a company that profits from GPU scarcity buy a company whose product reduces GPU demand? The answer is that distillation moves the workload, it does not eliminate it. You still train the small model. You still run RL inside sandboxes. You still generate traces. The cost-optimization demand gets converted into a new compute consumption point. That is the real logic of the OpenPipe deal, and the announcement never says it out loud.
The Registry Is the Only Real Anti-Lock-In Design
CoreWeave markets Forge as cloud-agnostic. I want to be very careful here, because I have watched this movie in crypto before. Every "chain-agnostic" bridge turned out to have a canonical side.
The on-cluster version of Sandboxes must run inside CoreWeave's Kubernetes Service. The "no incremental cost" benefit applies only to existing CoreWeave customers. So cloud-agnostic is the marketing layer, and land-and-expand compute binding is the actual layer. That is not a scandal. It is a standard enterprise playbook. But calling it cloud-agnostic is imprecise at best.
The one genuinely anti-lock-in component is Registry, described as an "open, portable format." If that format is OCI Artifacts, a competitor platform can import lineage data without loss, and the lock-in is soft. If it is a proprietary spec, the lock-in is hard and the "open" label is decoration. This single detail determines whether Forge is a platform or a trap, and the announcement does not specify it. That omission is itself information.
The Pricing Page Is the Strategy Document
Now to the number that started this piece. $60 per month for up to fifty seats. Under $1.20 per head. Free for individuals.
This is not a product launch. It is a defensive-suicide pricing move, and I mean that as a technical term. CoreWeave is willing to earn nothing on the control plane to prevent independent MLOps vendors from building stickiness inside its own GPU customer base. LangSmith, Langfuse, Braintrust, Arize โ these vendors' entire differentiation is cloud neutrality. CoreWeave's counter is free-or-near-free plus GPU adjacency. For a customer already renting CoreWeave GPUs, the switching cost collapses to zero and the benefit is immediate. The independent vendors face a binary fate: get acquired, or get strangled by a bundled free feature. W&B and OpenPipe are the receipts for the first outcome.
And CoreWeave is not alone in this. DigitalOcean priced Agent Droplets at $50 per month, within days and within dollars of CoreWeave's $60. Two independent vendors converging on the same price point inside the same week is not a coincidence. It is a commodity signal. When the control plane has no pricing power, the money has already migrated to the compute layer. This is the central commercial fact of the agent infrastructure market in 2025, and it is exactly the opposite of what the crypto AI narrative wants to believe.
There is an early-customer quality signal worth noting. Canva โ a scaled design SaaS with real production AI workloads โ appears as an early Forge customer. That is more strategically meaningful than a celebrity or a demo-stage startup, because it suggests Forge has reached enterprises with genuine scale. But two disclosed customers is not a sample. It is an anecdote. I would need a dozen to call it penetration.
The Competitive Matrix Nobody Published
The coverage of Forge that I have seen consistently frames DigitalOcean as the primary rival. That framing is convenient and wrong. DigitalOcean is a lightweight general-purpose cloud with weak GPU supply certainty. The real competitors are the hyperscalers, and they arrived first.
AWS shipped Bedrock AgentCore with Runtime, Memory, Gateway, Identity, and Observability. Azure shipped AI Foundry Agent Service. Google shipped Vertex AI Agent Builder. These platforms offer heavy functional overlap with Forge โ and they carry advantages CoreWeave cannot match: mature ecosystems, deep enterprise relationships, and in-house frontier models.
| Dimension | CoreWeave Forge | AWS Bedrock AgentCore | Azure AI Foundry | DigitalOcean | Independent MLOps | |---|---|---|---|---|---| | Positioning | GPU-native agent platform | General cloud agent platform | General cloud agent platform | Lightweight agent platform | Point tool | | Sandbox isolation | Kata VM (strong) | Firecracker / custom | Container / VM | Firecracker microVM | Varied | | Observability | Agent Lens | AgentCore Observability | Foundry Tracing | Basic | Strong (core product) | | Own models | None | Nova / Claude | Phi / GPT | None | None | | Pricing | $0 / $60 / enterprise | Usage-based | Usage-based | $50/mo | ~$39/seat | | GPU supply certainty | Extremely high | Medium (quota-bound) | Medium | Low | N/A | | Ecosystem maturity | Weak | Strongest | Strong | Medium | Weak-medium | | Enterprise relationships | Weak (concentrated) | Strongest | Strong | Medium | Weak |
The matrix is not kind to Forge on the axes that matter over a five-year horizon. Its one overwhelming advantage is GPU supply certainty โ priority NVIDIA allocation, early Vera CPU access, a $6.3 billion backstop. But that advantage is also a dependency. If NVIDIA's strategic intent shifts, or if GPU scarcity eases, Forge's moat evaporates. A moat built on the scarcity of a good your supplier controls is not a moat. It is a lease.
I have been doing on-chain verification long enough to distrust any advantage that cannot be independently confirmed by a second party. CoreWeave's GPU advantage is real and confirmable. Its platform advantage is asserted and unconfirmed. Those are very different things, and the coverage is treating them as one.
The Contrarian Angle: DePIN Compute Is Fighting on the Wrong Axis
Here is what the crypto-native reader should extract from Forge, and it is not a bullish signal.
The decentralized compute narrative โ Akash, Render, io.net, Aethir, Gensyn โ has spent years arguing that it can undercut centralized GPU rental on price. That argument was plausible when centralized providers charged a premium for scarcity. It becomes much weaker when the largest centralized GPU provider starts giving away the software layer for free and monetizing on utilization. The centralized player can subsidize the entire developer experience. A DePIN network that charges for orchestration cannot.
But the deeper problem is what DePIN is not building. CoreWeave built Forge around the full lifecycle: run, observe, curate, improve, evaluate. DePIN compute networks, almost without exception, built around raw capacity: here is a GPU, here is an hourly rate, here is a token. They solved the commodity layer and left the control plane to the centralized incumbents. That is a structural error, and Forge is the proof.
I want to be specific about where this bites. Take sandbox isolation. Forge chose Kata for strong isolation โ a decision that signals it is targeting regulated, compliance-bound enterprises. DePIN compute networks rarely offer VM-level isolation with per-sandbox kernels; they offer containers, or worse, unverified execution. An enterprise evaluating agent infrastructure will not put production workloads on unverified containers to save 15% on hourly GPU rates. The savings do not survive the compliance review.
Take observability. Agent Lens records traces. On a DePIN network, who records the trace, where does it live, and who can read it? Most networks have no coherent answer. This is the same failure mode I watched during the 2021 NFT metadata investigation, when my scraper found that 15% of popular collections were pointing at centralized servers while marketing "on-chain" permanence. The decentralization was cosmetic. The same cosmetic decentralization is being sold in compute today, and the market has not yet run the scraper.
Take pricing. DePIN compute price discovery depends on oracle feeds โ spot rates, utilization, token emissions. Oracle feed latency is the quiet failure mode here. When a compute market's price is set by an oracle that updates slower than the underlying GPU spot market moves, arbitrageurs extract the difference and the network subsidizes them. I have watched this pattern in DeFi for years. It is not a hypothetical. It is a structural cost that DePIN compute networks rarely price in, and it is exactly the kind of detail that never appears in a token launch thread.
And then there is the open-source absorption problem. Forge is built on Kata Containers and, reportedly, integrates Marimo, an open-source reactive notebook. Neither is a CoreWeave acquisition as far as public records show โ Marimo in particular looks like integration, not ownership, and the "acquisition" framing I have seen circulating is imprecise. But the pattern holds: open-source infrastructure gets commoditized into a commercial platform, the community that built it captures none of the upside, and the vendor captures all of it. This is the same public-goods funding failure that every grant committee in crypto has reproduced. The one mechanism that has consistently worked in this space is RetroPGF โ retroactive funding of already-delivered public goods โ precisely because it pays for proven work instead of promising future work through committee discretion. CoreWeave absorbing Kata and Marimo into a closed platform is the market solving a public-goods problem with private capture. It works for CoreWeave. It does not work for the maintainers, and it does not work for the ecosystem that depends on them.
So the honest framing is this: Forge is not primarily a threat to DigitalOcean. It is a demonstration that the centralized compute layer can absorb the entire decentralized compute value proposition faster than DePIN networks can build it. The hyperscalers did it with ecosystems. CoreWeave is doing it with GPU adjacency and free software. DePIN compute is left arguing about hardware costs while the control plane is being given away.
None of this means DePIN compute is dead. It means the differentiation has to move. The networks that survive will be the ones that offer something the centralized layer structurally cannot โ verifiable execution with cryptographic proofs, permissionless capacity that no backstop agreement can constrain, or data sovereignty that survives a terms-of-service review. Raw cheap GPUs is not that thing. It never was, and Forge just made the point for them.
Takeaway
The next twelve months will tell us whether the control plane for AI agents is genuinely worthless or merely being priced as worthless by one well-capitalized landlord. Watch three things. First, whether CoreWeave ever discloses a GPU conversion rate โ the average compute revenue generated per Forge seat. Without that number, the $60 price is a marketing gesture, not a business model. Second, whether the Registry's "open format" turns out to be OCI Artifacts or a proprietary spec; that single detail determines whether Forge is a platform or a trap. Third, and most important for anyone holding DePIN compute tokens, whether any decentralized network ships a control plane that competes on the lifecycle rather than the hourly rate.
If none of them do, then the GPU landlords will keep giving away software for a dollar twenty a seat, and they will keep winning. The question every crypto compute founder should be answering this quarter is simple: what are you selling that CoreWeave cannot give away for free? If the answer is capacity, you have already lost.