# Open Source Is Dead. Long Live Open Source.

Saoud Rizwan

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/CoEIs6Xm8m8-open-source-is-dead-long-live-open) · [Video](https://www.youtube.com/watch?v=CoEIs6Xm8m8)

## Open Source Is Dead. Long Live Open Source.

Cheap code strains the trust behind open-source contributions, while expensive inference makes open weights increasingly valuable as a foundation for competition.

### Why trust mattered when every request cost money

[0:18](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=18s)

Before prompt caching and coding-agent subscriptions, an agent that could finish a developer’s work could also generate an alarming bill. Saoud Rizwan recalls some early [Cline](https://cline.bot/) users spending hundreds of dollars a day on individual API requests. He describes Cline as the first coding agent, preceding Claude Max and Codex subscriptions. For those users, watching an LLM perform work end to end was compelling enough to keep paying.

Open source made that expense easier to trust. Developers could inspect the code and connect their own API provider, seeing what the agent did and where their money went. The community also shaped the product: Rizwan credits conversations with contributors for custom rules and plan mode, features he says Cline introduced first. Openness supported both adoption and product development because users could understand the system and help improve it.

### When contributions stop building contributors

[1:28](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=88s)

The same technology now puts pressure on the relationships that helped Cline grow. Generating a pull request, issue or security report is cheap; determining whether it deserves attention still consumes a maintainer’s judgment. Rizwan describes a community increasingly skeptical of whether strangers have used AI responsibly. Projects that depend on third-party contributions inherit that uncertainty with every submission.

Zig, the language used to build Bun, provides the sharpest example. The displayed code of conduct highlights a strict prohibition on LLM and AI use; Rizwan describes its scope as pull requests, issues and comments. His explanation is that Zig values developing contributors more than acquiring individual contributions. Reviewing a patch is an investment in someone who can eventually become a trusted collaborator. If AI assistance obscures what that person understands, the review no longer serves the same purpose.

![Zig Code of Conduct page with a red outline around “Strict No LLM / No AI Policy.”](https://ai.engineer/api/talk-essay-media/f1ccac825d4da868dc4b3e6dc823ccd75e9a53a2ac1b22a14f830ab62899f400)

Zig’s code of conduct highlights a “Strict No LLM / No AI Policy.”

Other projects have restricted the incoming work itself. Rizwan cites curl’s flood of AI-generated security reports and the possibility of ending its bug bounty. That account describes an earlier stage: Daniel Stenberg’s subsequent update says the bounty ended in February 2026. The [tldraw contribution policy](https://github.com/tldraw/tldraw/issues/7695) temporarily closes external pull requests automatically, with selective reopening, while retaining issues, bug reports and discussions. GitHub also introduced controls to disable pull requests entirely or restrict their creation to collaborators with write access. These are different degrees of restriction, rather than a single decision to abandon all community participation.

**The part of open source under pressure is the cultivation of trust.** Rizwan’s claim that open source is dying concerns the economics of maintaining those relationships when producing software becomes cheaper but evaluating unfamiliar contributors remains costly. It does not mean that published source code has stopped being useful.

### One compromised dependency reaches downstream

[3:56](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=236s)

Trust also extends through the software supply chain. An application can depend on a long sequence of maintainers, build tools and publishing systems; compromising one of them can deliver malicious code to downstream users. Rizwan uses the LiteLLM Python package to make this exposure concrete, describing roughly 3.5 million daily downloads and a malicious-release window of about three hours. Download volume indicates the package’s reach, not the number of infected users.

The attack crossed several trust boundaries. The documented entry point was compromised Trivy tooling in LiteLLM’s CI pipeline, refining Rizwan’s shorthand description of a GitHub app compromise. Stolen PyPI publishing credentials allowed attackers to distribute malicious releases. Once installed, the payload collected API keys, SSH keys and cryptocurrency credentials, and attempted to establish persistent access for remote command execution. The diagram follows that path from a security tool through publishing access to users’ machines.

![Six connected steps show compromised Trivy, LiteLLM’s CI/CD pipeline, secret exfiltration, stolen publishing access, malicious packages, and theft of user keys with backdoor installation.](https://ai.engineer/api/talk-essay-media/be38da1f373b0f814dae7b263425f62d885a46417de31bb47283f626831be0eb)

A diagram traces the compromise from Trivy through LiteLLM to credential theft and a backdoor.

Discovery depended on a defect in the malware. Rizwan describes a Cursor crash involving MCP; the discoverer’s account more precisely identifies a Cursor MCP plugin that pulled LiteLLM as a transitive dependency, followed by recursive Python startup that crashed the machine. A researcher investigated the failure and uncovered the payload. For enterprise users operating internal LLM gateways, the potential exposure extended beyond a disposable development environment to credentials for other systems. Rizwan characterizes the quick discovery as luck, rather than evidence that the dependency chain reliably caught the attack.

### The value that survives: using and building on the work

[5:22](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=322s)

The ability to use someone else’s work and build on it remains valuable even as contribution channels become harder to maintain. That freedom comes through the applicable license; open source does not necessarily mean public domain. For model weights, openness has an additional economic consequence: more organizations can participate in serving and optimizing the same model. The reason to care becomes clearer when inference bills reach organizational scale.

Rizwan cites an anonymous CFO account alleging that an unnamed company spent $500 million on **Claude** in one month after failing to set usage limits for thousands of employees. The anonymous account cannot establish a typical enterprise bill, but its failure mode is straightforward: broad access without spending controls can turn individual usage into an organization-wide liability.

In a separate report attributed to Uber’s CTO, Rizwan cites 95% engineer adoption, 70% of committed code coming from Claude, monthly spending of up to $2,000 per user, and exhaustion of the entire 2026 budget in four months. These are reported figures, not independently established measurements here. Together they illustrate the transition from experimenting with an assistant to budgeting for infrastructure that engineers use throughout their work.

### Subscription value can become application dependence

[6:27](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=387s)

Subscription pricing can conceal how much inference a workflow consumes at retail API rates. Rizwan describes SemiAnalysis experiments that ran long-horizon coding tasks in Claude Code and Codex until their weekly subscription limits were exhausted. He reports the following API-equivalent values:

| Subscription | Reported API-equivalent usage |
|---|---:|
| $200 Claude plan | About $8,000 |
| $200 Codex plan | About $14,000 |

These figures describe retail API-equivalent usage under that workload, not measured provider costs or losses. The account does not resolve the measurement period, extrapolation or caching assumptions behind the totals.

Rizwan interprets this pricing as a subsidy intended to establish dependence. Agents in CI, background agents and looping agents introduce more places to spend tokens, and each new workflow can make a particular application harder to replace. His predicted next step is higher prices once developers cannot comfortably work without the tools; that prediction is separate from the subscription experiment.

The audience gets a more immediate test of dependence: who stops working during a Claude or **GPT** outage? Rizwan includes himself among those affected. In his reading, Anthropic and OpenAI’s investment in applications gives them a potential moat even if competing models converge in capability. The durable asset would be the workflow surrounding the model, rather than exclusive access to intelligence.

### Put more intelligence into the development system

[8:06](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=486s)

That application moat may prove weaker than expected if developers can obtain adequate results for less money elsewhere. Rizwan acknowledges that many open-weight contenders, including models developed in China, have historically trailed the leading closed models. His proposed inflection point is practical sufficiency: a business does not need the strongest available model for every task, and better CLI features may not outweigh a large difference in operating cost.

**Model capability is only one part of an agent’s result.** The surrounding system determines what information the agent sees, what it can do and what counts as completion:

- **Context and tools:** Give the agent the project information and operations needed to solve the task.
- **Skills and rules:** Encode project conventions and reusable procedures instead of relying on the model to infer them each time.
- **Verification and quality gates:** Require checks of the resulting work before accepting completion.

Rizwan argues that this infrastructure can help a less capable model produce results similar to a stronger one.

The tradeoff may be more tokens: extra investigation, tool use and verification can compensate for weaker initial judgment. If those tokens are cheaper, the additional work can still cost less. Placing expectations and checks in the system also reduces reliance on each developer remembering to use the model responsibly. The goal is to make successful completion a property of the development process, not merely the agent’s confidence.

### The same bug, different definitions of done

[9:28](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=568s)

Skeptical of benchmarks placing **GLM** ahead of **Opus**, the Cline team tried both on a real bug in the [Cline repository](https://github.com/cline/cline). Rizwan presents this as a single-task anecdote. Both models addressed the issue, but their cost and completion behavior differed.

| Dimension | GLM | Opus |
|---|---|---|
| Tokens used | Twice as many as Opus | Half as many as GLM |
| Cost | About half of Opus | About twice GLM |
| Completion speed | Slower | Faster |
| Tool calls | Twice as many as Opus | Half as many as GLM |
| Verification and cleanup | Removed dead code; checked compilation | Left type errors; broke the production build |

Rizwan reports these results for the same Cline repository bug. The completed slide labels the runs **GLM-5.2** and **Opus 4.8**, with costs of $0.41 and $0.81 respectively. Its build outcomes make the practical distinction visible: addressing the reported issue was not sufficient if the resulting project could not build.

![Slide titled “Fixing the same bug. Half the cost.” Purple GLM-5.2 bar reads $0.41 and “builds clean”; orange Opus 4.8 bar reads $0.81 and “breaks the build.”](https://ai.engineer/api/talk-essay-media/f2aef139dcaf406569b1113caf07512a38743b2c8fb247e5b75d0941f55f1aa1)

The bug-fix comparison shows GLM-5.2 at $0.41 with a clean build and Opus 4.8 at $0.81 with a broken build.

The team inferred that GLM had been trained to spend more tokens verifying its output. The observed behavior supports the narrower conclusion that this run did more verification; it does not establish the training cause. For the workflow, however, the economic point remains useful: a slower run with more tool calls can be preferable when it costs less and leaves a working build.

### An internal gateway makes switching practical

[10:18](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=618s)

The same economics can shape enterprise defaults. Rizwan cites Coinbase CEO Brian Armstrong as saying that Coinbase’s internal LLM gateway defaulted to GLM and **Kimi**, nearly halving AI spending while token usage continued to grow. In that account, lower spending did not require using AI less; it came from changing which models served the work.

An internal gateway gives the organization a place to make those routing decisions without standardizing every workflow on one model vendor’s application. Rizwan expects businesses to build internal tooling around dollar efficiency, even when that means forgoing a newly released Claude Code feature. The relevant comparison becomes the cost of getting the organization’s work done through its own system.

### Open Compute changed what suppliers manufactured

[10:57](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=657s)

Rizwan’s historical analogy is the Open Compute Project. In his account, Facebook developed data-center infrastructure around 2011, but Amazon and Google had already built enough of their own that the work no longer provided a distinctive competitive advantage. Facebook released designs for servers, data centers, networking and cooling hardware, including schematics and CAD files. Publishing the designs let other buyers and manufacturers build around the same work.

The mechanism is a change in production economics:

| Before shared designs | With a shared standard |
|---|---|
| Buyers specify proprietary hardware | Buyers order common designs |
| Manufacturers make small custom runs | Manufacturers make larger standardized runs |
| Custom components support vendor premiums | Interchangeable supply constrains premiums |

As demand converged, manufacturers could spread their effort across larger production volumes. Rizwan credits the resulting market with lowering costs for the wider industry and saving Facebook billions of dollars. Giving away the design could therefore reduce the originator’s own future purchasing costs.

The lesson he draws is that the industry can standardize on something it can build upon even when it is not the strongest option in every technical dimension. For open weights, the corresponding opportunity is to let many providers invest in making the same models cheaper and easier to serve.

### More capacity, more ways to lower inference cost

[12:34](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=754s)

Rizwan cites forecasts of nearly $3 trillion in AI infrastructure spending and more than 100 GW of new data-center capacity by 2030, roughly doubling global capacity relative to the talk’s present. These are projections rather than completed investments. His expectation is that the build-out will give inference providers more infrastructure on which to compete.

Providers such as Baseten and Fireworks can compete to serve open weights more efficiently through several complementary mechanisms:

- **Dedicated hardware:** Match the serving environment to a known workload.
- **Caching:** Reuse eligible computation rather than repeating it.
- **Batching at volume:** Combine work to improve hardware utilization.
- **Inference-specialized silicon:** Use hardware designed around inference requirements.

Open weights allow multiple hosts to pursue these improvements for the same model, exposing more of the efficiency gains to competitive pricing.

A further forecast cited by Rizwan puts inference for a one-trillion-parameter LLM at 90% lower cost by 2030 than at the time of the talk. The originating assumptions are not specified, so the forecast supplies a direction and scale for his argument rather than a dependable price schedule.

### Cloud competition moved beyond raw compute

[13:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=800s)

Cloud pricing supplies a second precedent. Rizwan recalls Google’s March 2014 GCP Live announcement cutting compute prices by 32% and storage prices by 68%, followed within days by what he describes as AWS’s 42nd price cut. In his account, competition from 2015 onward increasingly shifted toward databases and serverless services as basic compute and storage became commodities.

He expects open-weight inference to follow a similar path: hosts compete aggressively on serving efficiency, customers adopt the models that make economic sense, and differentiation moves elsewhere. His prediction is that foreign open-weight models will gain broad adoption as closed-lab API premiums become harder to justify for much knowledge work. That outcome depends on adequate task performance as well as price; cheaper tokens alone do not complete the job.

### Release weights without giving away the research lead

[14:22](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=862s)

The strategic concern follows from standardization. If the infrastructure now being built converges on foreign open-weight models, later marginal improvements in GPT, Claude or **Gemini** may provide too little reason to switch. Rizwan urges American labs to take open weights more seriously while adoption and developer mindshare remain contestable.

He also connects leadership to the values of the organizations and countries developing the technology. He worries that losing Anthropic and OpenAI’s influence would diminish the role of their safety-research investment. This is his rationale for wanting those labs to remain central to the ecosystem, rather than a technical finding about the safety of particular models.

His proposed distinction is between **releasing model weights** and **releasing the research that produces the next model**. He regards the latter as the source of a continuing lead. Weights can make it easier to extract traces and train imitation models, but he argues that this does not by itself give a competitor the ability to leapfrog the originator. In his proposal, broader hosting competition, adoption and customer value would help preserve leadership by making the models useful foundations for the rest of the industry.

### Cline’s bet on competing providers

[16:01](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=961s)

Cline’s closing announcement applies that thesis to its own offering. Rizwan says the team launched an open-weight subscription earlier that week, using volume discounts and partnerships with inference hosts to offer lower costs than direct API usage. The slide names the offering ClinePass. He says Cline plans to increase usage quotas as serving costs fall, and points to GLM and **DeepSeek** as models developers can try through it. No numerical subscription price or quota is given in the presentation.

![ClinePass slide with the text “The best open weights subscription,” cline.bot/pass, and surrounding DeepSeek, Qwen, Kimi, GLM, MiMo, and Minimax labels.](https://ai.engineer/api/talk-essay-media/7b575ce7937f5ac7bb7668184497dedb4411ead9b97a3f6ea0221baebf2bbffe)

ClinePass presents an open weights subscription at cline.bot/pass.

The subscription sits alongside Cline’s open-source agent and bring-your-own-key access to other providers. Rizwan lists CLI, VS Code and JetBrains interfaces; the current Cline site labels JetBrains support Early Access. Provider choice remains part of the product rather than requiring every user to adopt the subscription.

Cline also continues adding newly released models so developers can compare what improved capability means for their own work, including open-weight models outside Claude and ChatGPT subscriptions. That is the concrete version of the closing economic argument: keep the development environment usable across providers, and let models compete on the results and value they deliver inside it.

## Resources

- [Cline source repository](https://github.com/cline/cline) — Source repository for Cline's coding agent, including its IDE, CLI and SDK surfaces.
- [tldraw's contributions policy](https://github.com/tldraw/tldraw/issues/7695) — The maintainer's announcement of temporary external pull-request restrictions and the contribution channels that remain open.
- [curl's changing security-report workload](https://daniel.haxx.se/blog/2026/04/22/high-quality-chaos/) — Daniel Stenberg describes report quality and maintainer workload after curl ended its bug bounty.
- [The original LiteLLM compromise disclosure](https://futuresearch.ai/blog/litellm-pypi-supply-chain-attack/) — The discoverer's account of the malicious PyPI releases, the crash that exposed them, and the credential-stealing payload.
- [LiteLLM attack chain and timeline](https://snyk.io/blog/poisoned-security-scanner-backdooring-litellm/) — Security analysis connecting the LiteLLM releases to compromised Trivy tooling and describing the exposure window.

## Slides

### Open source is dead, long live open source

Saoud Rizwan credits Cline’s open source code with helping developers inspect and trust it, and its community with shaping features such as custom rules and plan mode.

[0:19](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=19s) [Title slide introducing Saoud Rizwan, founder of Cline.](https://assets.ai.engineer/sites/aiecode2025/sha256/1020aff5c0afac1ff50eee985c9bc8b2189e81a0c168e693909c68f773265fa3.webp)

Open source is dead, long live open source

Saoud Rizwan @sdrzn
Founder @cline

### Zig’s no-AI policy

Zig’s policy bans LLM-generated content and several uses of AI in specified spaces. Rizwan says the project values growing trusted contributors and sees AI assistance as undermining that goal.

[2:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=140s) [Zig code of conduct page with its strict no-LLM and no-AI policy highlighted.](https://assets.ai.engineer/sites/aiecode2025/sha256/2dad32263834279b5898ce8686e01663ecaec1c0b92c87bf38c0a71781c388b6.webp)

Code of Conduct

The Zig community is decentralized. Anyone is free to start and maintain their own community, which is not subject to these rules.

This document contains the rules that govern these spaces only:
• The ziglang organization on Codeberg
• #zig IRC channel on Libera.chat
• Zig project development Zulip chat

The rules here are strict. This space is for focused, on topic, technical work on the Zig project only. It is everyone's responsibility to maintain a positive environment, especially when disagreements occur.

Strict No LLM / No AI Policy

No LLM-generated content, whether it be code or prose.

No paraphrasing LLM-generated content.

No LLMs for editing, including fixing spelling or grammatical errors.

No LLMs for translation. English is encouraged, but not required. You are welcome to post in your native language and rely on others to have their own translation tools of choice to interpret your words.

No LLMs for brainstorming and then sharing the results of that brainstorming, even if you create the prose. If you use a chatbot to give you advice on a comment on the issue tracker, that comment is unwelcome.

No LLMs for finding bugs.

No talking about use of chatbot/LLM services.

Profession by Isaac Asimov

Interface text:
ZIG
Download  Learn  News  Zig Software Foundation  Devlog  Source  Join a Community
← Back to Home

### curl’s response to AI-generated reports

Daniel Stenberg says curl is being overwhelmed by AI-generated security reports. Reporters must disclose AI use, and those submitting reports deemed “AI slop” are banned immediately.

[2:59](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=179s) [Daniel Stenberg’s post highlights a policy to ban reporters submitting reports deemed AI slop.](https://assets.ai.engineer/sites/aiecode2025/sha256/6ab4be31efc9f17c35361e7696052a53125fb508d7dc7d1ea96a1883d539c91d.webp)

Daniel Stenberg · 3rd+
curl CEO. Code Emitting Organism
1yr · Edited

That's it. I've had it. I'm putting my foot down on this craziness.

1. Every reporter submitting security reports on #Hackerone for #curl now needs to answer this question:

"Did you use an AI to find the problem or generate this submission?"

(and if they do select it, they can expect a stream of proof of actual intelligence follow-up questions)

2. We now ban every reporter INSTANTLY who submits reports we deem AI slop.

A threshold has been reached. We are effectively being DDoSed. If we could, we would charge them for this waste of our time.

We still have not seen a single valid security report done with AI help.

Interface text:
+ Follow
6,718
254
640
Add a comment...
GIF

### tldraw contribution policy

tldraw temporarily closes pull requests from external contributors after an increase in AI-generated submissions. Maintainers may reopen requests under consideration and continue to welcome issues and discussions.

[3:19](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=199s) [A tldraw GitHub issue highlights automatically closing pull requests from external contributors.](https://assets.ai.engineer/sites/aiecode2025/sha256/c3758266ef3ee97965bc24687b418ee658419a36a35da728862e221d894a0e92.webp)

Contributions policy #7695

Hey all, update on the tldraw policy with regard to contributions.

For the good of the project, we're going to begin automatically closing pull requests from external contributors. We will of course continue to welcome issues, bug reports, and discussions. This is a temporary policy until GitHub provides better tools for managing contributions.

Getting ahead of it

Like many other open-source projects on GitHub, we've recently seen a significant increase in contributions generated entirely by AI tools. While some of these pull requests are formally correct, most suffer from incomplete or misleading context, misunderstanding of the codebase, and little to no follow-up engagement from their authors.

An open pull request represents a commitment from maintainers: that the contribution will be reviewed carefully and considered seriously for inclusion. For that commitment to remain meaningful, we need to be more selective. For now, that means closing first and selectively re-opening pull requests that are actually under consideration, with the expectation that most unsolicited submissions will not be reviewed.

Weird year coming

I first made the tldraw repository public in 2021, and I've been proud to invite public contribution to its code and have been happy to have many improving pull requests over the years. I'm sorry to shut that down, however I sincerely believe this decision is in the best interests of the project, our code, and our community. With luck, GitHub will soon roll out management features that let us open things back up.

This is going to be a weird year for programmers and open source especially. For now, whether you've contributed before, are interested in contributing in the future, or just are a friend of the project: thank you and please hang on while we all figure this stuff out.

Interface text:
tldraw / tldraw
Type / to search
Code · Issues 365 · Pull requests 73 · Agents · Actions · Projects · Security and quality · Insights
Open · Task
steveruizok opened on Jan 15 · Collaborator
Reactions: 👍 185 · 👎 4 · 😄 2 · 😕 1 · ❤️ 88

### Pull request permissions

GitHub’s repository settings can restrict pull request creation to collaborators, while keeping existing pull requests readable to everyone who can see the repository.

[3:26](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=206s) [GitHub repository settings show pull request creation restricted to collaborators.](https://assets.ai.engineer/sites/aiecode2025/sha256/11dd8cc5dd9838fd787d9802f0b8a610157898cf1de4e2a199203092931d190b.webp)

Discussions is the space for your community to have conversations, ask questions and post answers without opening issues.

Get started with Discussions
Engage your community by having discussions right in your repository, where your community already lives
Set up discussions

Projects
Projects on GitHub are created at the repository owner's level (organization or user) and can be linked to a repository’s Projects tab. Projects are suitable for cross-repository development efforts such as feature work, complex product roadmaps or even issue triage.

Pull requests
Pull requests allow others to suggest changes to your repository.

Pull request permissions
Creation allowed by: Collaborators only
If restricted, pull requests will still be readable by everyone who can see this repository.

Interface text:
Code  Issues  Pull requests  Agents  Discussions  Actions  Projects  Wiki  Security  Insights  Settings
General
Access
Collaborators
Moderation options
Code and automation
Branches
Tags
Rules
Actions
Models
Webhooks

### LiteLLM supply-chain attack

A compromised Trivy scanner exposed LiteLLM’s publishing credentials, enabling malicious releases that stole keys and installed a backdoor on users’ systems.

[4:14](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=254s) [A diagram traces a compromised Trivy scanner through stolen publishing credentials to malicious LiteLLM releases and harm to users.](https://assets.ai.engineer/sites/aiecode2025/sha256/e478575ece796a02c70d5140d58297bee2687e0314d89d1986a0b867c5985802.webp)

Trivy scanner
(GitHub Actions marketplace) compromised

LiteLLM's
CI/CD pipeline runs compromised Trivy

Trivy exfiltrates
CI/CD environment secrets

Attacker obtains
PyPI publish token + GitHub PAT

PyPI token publishes
malicious 1.82.7 & 1.82.8
GitHub PAT defaces the maintainer profile & repos

END USER INSTALLS
Steals SSH keys
API keys · crypto keys
installs a backdoor
remote command execution

cline.bot

### Report on Claude spending

An article reports that an unnamed company spent $500 million on Claude in one month, highlighting the absence of employee usage limits as the cause.

[5:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=320s) [Article reports that an unnamed company spent $500 million on Claude in one month, highlighting the absence of employee usage limits.](https://assets.ai.engineer/sites/aiecode2025/sha256/21370bfa8f9355b32775dddc0c06925f0834387473b681793f75f791a2cb7f0e.webp)

Gadget Review_

Company Blew $500M On Claude AI In One Month Due To No Usage Limit On Licenses For Employees

Nikshep Myle
May 29, 2026 · 2 min read

A mysterious enterprise just torched $500 million in a single month on Anthropic’s Claude AI platform, according to Axios reporting. The culprit? No usage limits on employee licenses, turning what should have been controlled experimentation into a financial bloodbath that makes your surprise Netflix subscription charges look quaint.

You've probably experienced the sting of unexpected cloud bills before—maybe a few hundred dollars when that side project accidentally left servers running. Scale that feeling up by several million times. This anonymous enterprise learned the hard way that token-based AI pricing without guardrails transforms helpful productivity tools into budget-devouring monsters faster than a TikTok algorithm learns your guilty pleasures.

Interface text:
ANTH.P...
6

### Uber AI budget report

Uber exhausted its 2026 AI budget by April as Claude Code spread across roughly 5,000 engineers, the article reports. The speaker says monthly spending reached up to $2,000 per user.

[6:04](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=364s) [Article reports that Uber exhausted its 2026 AI budget by April as Claude Code spread among its engineers.](https://assets.ai.engineer/sites/aiecode2025/sha256/9ce8de4fc0367fc6148967c622e4cac0ae26b87736d4a3a57b13847a8320507b.webp)

EDITORS’ PICK | INNOVATION > CLOUD

Uber Burns Its 2026 AI Budget In Four Months On Claude Code

By Janakiram MSV, Senior Contributor. I cover emerging technologies with ...
Published May 17, 2026, 11:18pm EDT, Updated May 27, 2026, 12:17pm EDT

Taxi
PIXABAY

Uber exhausted its entire 2026 artificial intelligence budget by April, four months into the calendar year, after Anthropic's Claude Code spread across roughly 5,000 engineers faster than the company’s finance models had anticipated.

Chief Technology Officer Praveen Neppalli Naga confirmed the overrun to The Information, saying the company was back to the drawing board on its assumptions.

Uber’s total research and development spend reached $3.4 billion in 2025, up 9% year over year, which makes the budget collapse less about scale and more about a pricing model that enterprise finance teams have not learned how to manage.

The disclosure landed alongside a structural shift from Anthropic itself. On May 13, the company announced that paying Claude subscribers would soon face a separate monthly credit meter for agent tools and third-party harnesses, billed at full application programming interface rates starting June 15. Read together, the two events describe a single problem. Token-based consumption pricing does not behave like the software line items chief financial officers know how to model, and the gap between what engineers consume and what finance teams expect is no longer hypothetical.

Interface text:
Follow Author

### Subscription prices and estimated maximum spend

In SemiAnalysis tests, long-horizon coding tasks exhausted weekly limits. The $200/month Claude and Codex plans yielded about $8,000 and $14,000/month in API-equivalent usage, respectively.

[6:26](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=386s) [Table compares monthly plan prices with approximate maximum possible spend for Claude and ChatGPT plans.](https://assets.ai.engineer/sites/aiecode2025/sha256/b515ebf3fb190b7f7e69630b6a4d3a6b18efaceb72522b553b624046dd2bb31c.webp)

PLAN | PLAN PRICE | MAX POSSIBLE SPEND (APPROX.)

ANTHROPIC — CLAUDE
claude-pro | $20/mo | $400/mo
claude-max-5x | $100/mo | $2,000/mo
claude-max-20x | $200/mo | $8,000/mo

OPENAI — CHATGPT / CODEX
chatgpt-plus | $20/mo | $700/mo
chatgpt-pro-5x | $100/mo | $3,500/mo
chatgpt-pro-20x | $200/mo | $14,000/mo

### Model infrastructure and quality gates

Context, tools and verification surround the model. The speaker argues these guardrails can help a less capable model produce similar results, though it may use more tokens.

[8:49](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=529s) [Concentric layers around a model identify context, tools, and verification and quality gates.](https://assets.ai.engineer/sites/aiecode2025/sha256/baf7eb071994f21922c47ee6294a7bbb87ea8ced52d478134d6d413b5a3c597c.webp)

VERIFICATION & QUALITY GATES
tests · type-check · build · code review · CI

TOOLS
MCP · terminal · file edits · browser

CONTEXT
rules · skills · memory · docs

Model

### Bug-fixing cost comparison

In one Cline bug fix, GLM-5.2 built clean for $0.41; Opus 4.8 cost $0.81 and broke the build. GLM used twice as many tokens, while Opus finished faster with half as many tool calls.

[9:26](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=566s) [A slide compares one bug fix: GLM-5.2 builds clean at $0.41, while Opus 4.8 breaks the build at $0.81.](https://assets.ai.engineer/sites/aiecode2025/sha256/a372eb8959af0100c232cee33024eba9a9643e75c1af756624dde60c679540dc.webp)

Fixing the same bug. Half the cost.

GLM-5.2
✓ builds clean
$0.41

Opus 4.8
× breaks the build
$0.81

### Model defaults and routing

Brian Armstrong describes experimenting with GLM 5.2 and Kimi 2.7 as cheaper defaults while routing prompts to models based on the job, cache hits and pricing.

[10:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=620s) [Brian Armstrong’s post discusses AI spending, routing and caching, highlighting an experiment with open-weight model defaults.](https://assets.ai.engineer/sites/aiecode2025/sha256/b4629561cc457a2b44ca7f3d169557ca675fd2ba9919c398671547969896a8ad.webp)

Brian Armstrong
@brian_armstrong

How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching.

Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work.

Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task.

### Meta’s open data center designs

Facebook opened designs from data centers and racks down to servers, switches, power and cooling. The speaker says shared designs enabled larger production runs and lowered industry costs.

[10:55](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=655s) [Diagram of open data center, rack, server, network switch, and power and cooling designs.](https://assets.ai.engineer/sites/aiecode2025/sha256/242609aa31f721c2419206e5ac4fccacf2bd9d58712bc0790343dfa3b0a1fd19.webp)

Meta open-sourced THEIR data center.

OPEN Compute Project ®

Data center design — OPEN

Rack & chassis — OPEN
power · cooling · the physical frame

Server & motherboard — OPEN
stripped-down compute, no vanity parts

Network switch — OPEN
"the wedge" — an open switch design

Power & cooling — OPEN
power shelf & thermal design

### Compute growth and cost projections

The chart projects U.S. data-center power rising from ~31 GW in June 2026 to ~95 GW in 2030. The speaker estimates that inference on a 1-trillion-parameter LLM will cost 90% less in 2030 than today.

[12:31](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=751s) [Chart projects U.S. data center power rising from about 31 to 95 GW by 2030 and indexed intelligence serving costs falling 90%.](https://assets.ai.engineer/sites/aiecode2025/sha256/0d4ca28d43de743503af9bf1c1bcd1350c595d00418a51316cce2a1da02493c9.webp)

Compute floods in. Cost collapses.

Legend:
U.S. data-center power (GW)
Cost to serve intelligence (indexed)

Chart labels: today; ~31 GW; ~95 GW; −90%
Timeline: Jun 2026; 2027; 2028; 2029; 2030

### Web hosting and model hosting

Model hosts use cost-cutting techniques that echo cloud hosting: continuous batching shares GPUs, specialized chips speed inference, caching avoids repeat work, and off-peak batches use cheaper capacity.

[13:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=800s) [Comparison pairs web hosting efficiencies with model hosting techniques, including batching, specialized chips, caching, and off-peak processing.](https://assets.ai.engineer/sites/aiecode2025/sha256/8c23f578f37fd491c95f59103cd9c36155857dd8747bcac96bd2225732a539ca.webp)

WEB HOSTING
AWS  GCP  Azure

MODEL HOSTING
Baseten  Fireworks  Together

Multi-tenancy
many tenants share one box
Continuous batching
many requests share one GPU

Custom silicon
Graviton · Nitro · TPU
Inference chips
Trainium · TPU · Groq · Cerebras

Edge caching & CDNs
don't re-serve the same bytes
KV & prompt caching
don't recompute the same tokens

Spot & preemptible
idle capacity, sold cheap
Batch & off-peak
async inference at half price

### Closed research and open weights

The speaker urges labs to release model weights while keeping research private, arguing that adoption and hosting competition can spread the models and improve customer value.

[14:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=860s) [Slide contrasts keeping training methods, data pipelines, and architecture insights closed with claimed adoption and cost benefits of open weights.](https://assets.ai.engineer/sites/aiecode2025/sha256/150b47aa2fb02e19ed8961abc5fe2c240e5584eff29a34a7c6eb22390be6d0d9.webp)

Closed Research

Training recipes & methods
the how, not the what

Data & RL pipelines
the real moat

Architecture insights
where the edge comes from

Open Weights

Can't be reverse-engineered
you give away the output, not the lead

Drives adoption & standardization
become the substrate the industry builds on

Reduced costs & better value
hosting competition drives price to the floor

### ClinePass open weights subscription

ClinePass offers an open-weights subscription. The speaker says volume discounts and inference-host partnerships lower costs versus direct API pricing, and plans to increase quotas as models become cheaper.

[15:59](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=959s) [ClinePass subscription slide surrounded by labels for several open-weight models.](https://assets.ai.engineer/sites/aiecode2025/sha256/3ce4817ce94c4c7b653f9b1dbbef87b2061ed53bdac058732fc58be4abf591e9.webp)

ClinePass
The best open weights subscription
cline.bot/pass

Qwen 3.7 Max
DeepSeek v4 pro
Kimi k2.7 code
GLM 5.2
MiMo v2.5
Minimax m3

## Full transcript

[0:00](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=0s) [outro jingle] Hi, I am Saoud, um, founder of Cline.

[0:18](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=18s) Uh, I started Cline as an open source project a few years ago. Um, some of you might know it as the first ever coding agent back before, um, the Claude Max subscription and the Codex subscriptions when people had to pay for each and every API request, which, uh, got extremely expensive.

[0:37](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=37s) Um, this was before prompt caching became a thing. And so there were people that paid hundreds of dollars a day, uh, using Cline. But for a lot of people, it was their first AGI moment.

[0:50](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=50s) It was the first time they saw, um, LLMs be able to do their jobs end to end, um, and they got hooked. Um, and I don't think Cline would have been as successful at as it, as it is if it wasn't open source, um, because it allowed these developers to inspect our code and trust it and, uh,

[1:08](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=68s) connect to any API, so they could be comfortable with spending so much money on it, um, and, uh, know that they weren't getting screwed over.

[1:17](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=77s) Um, and we were the first to add things like custom rules and, and plan mode, and a lot of that came from talking to and learning from this really incredible open source community we had around the project.

[1:28](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=88s) Um, and so, you know, having spent most of my life building open source, it's really heartbreaking to see just, like, the broader open source community, uh, wither and, and die over the last two years, uh, because of how AI has fundamentally changed everything about software development.

[1:45](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=105s) Um, GitHub is effectively, uh, an archive of slop PRs and issues and security reports, um, where the sense of community before has turned into this, like, deep skepticism and distrust, um, of each other's responsible use of these tools.

[2:03](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=123s) Um, because AI coding can be extremely dangerous to a project, and everyone's kind of had to learn that on the fly, but especially open source projects, uh, that rely on trusting third parties.

[2:15](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=135s) Um, and so I wanted to share some examples of how open source has been dealing with AI.

[2:22](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=142s) So this is the code of conduct for Zig, which is the language that powers Bun, um, and they essentially ban all use of AI. You can't use it on pull requests or issues or even comments.

[2:36](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=156s) Um, and the reason for this is that to them, the core Zig team, um, they value contributors more than they do the, the contributions. Um, and so the primary goal for reviewing PRs and things isn't to add new code, but it's to help grow new contributors who can become trusted, um, over time, and AI assistance completely breaks

[2:57](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=177s) that. Um, this is a, uh, a post from the CEO of curl who says that his project is effectively being DDoS-ed by AI-generated bug reports, and they're even considering, um, shutting down their bug bounty program for the first time in, in decades.

[3:17](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=197s) Um, and this is tldraw. Uh, they're automatically just closing all pull requests, whether they're AI-generated or not. Um, and it's gotten so bad that GitHub added a feature to disable third-party pull requests altogether, um, which is, which is really sad because pull requests were the thing that made GitHub what it is today, and we're probably gonna see

[3:41](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=221s) a lot of big open source projects, um, opt into this.

[3:45](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=225s) Um, and so when I say that open source is dead, I mean some parts of it, like the community, um, it's just not worth cultivating anymore, especially because building software is so cheap.

[3:56](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=236s) Um, and, uh, also the risk of supply chain attacks. I'm sure you've seen all the, all the reports of, of things getting compromised. It's become more dangerous than ever to depend on third-party software, uh, where it takes a single compromise in a massive chain of contributors to get pwned.

[4:14](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=254s) Um, so just as an example, LiteLLM, um, is a Python package. It gets, like, three and a half million downloads a day. They were compromised for three hours where attackers, um, used a GitHub app that they used, um, to steal their PyPI publishing tokens and, um, publish a compromised version of the package that would install a credential

[4:39](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=279s) harvester that would steal your API keys, your SSH keys, your crypto keys, um, and also install a backdoor that lets them do remote command execution. And the only reason this was even caught as quickly as it was, was just pure luck, um, 'cause the malware had a bug in it where it would cause Cursor to crash if

[4:59](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=299s) you ran the LiteLLM MCP server. And a security researcher noticed that, um, and was able to figure it out, but if this had been out any longer, it would have caused, like, catastrophic damage, especially because a lot of the people using LiteLLM are, like, the enterprise customers, um, and developers, um, that have their own internal gateways.

[5:22](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=322s) Um, but despite all of this, I believe there are some parts of open source that are sticking around and becoming more important, um, like allowing others to use your thing and, uh, freely in the public domain and build on top of.

[5:36](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=336s) And those parts about it are gonna become more important than ever, um, particularly with open weights models because of the economic impact. And so to help explain why, I wanna look at what's happening with inference spend right now.

[5:48](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=348s) So this is a report from, uh, an anonymous report from a CFO at an unnamed company where they accidentally spent five hundred million dollars on Claude in a single month because they didn't set the usage limits on their thousands of employees on their Anthropic dashboard.

[6:05](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=365s) Um, this is another report by Uber CTO where after they rolled Claude out to their organization, uh, 95% of their engineers were using it, seventy-fi- seven- 70% of their committed code came from Claude, and their monthly spend per user was up to $2,000, and they said they used their entire 2026 budget, um, in just four months.

[6:27](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=387s) And the crazy part is, is that the AI labs are losing money, too. Um, this is a chart from SemiAnalysis where they ran experiments with Claude Code and Codex subscriptions where they would give them long, um, horizon coding tasks until they exhausted their weekly limits.

[6:43](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=403s) And they found that a $200 plan for Claude would give them about $8,000 worth of API usage, um, and a $200 subscription to Codex would give them about $14,000 worth of API usage.

[6:56](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=416s) Um, so I think the strategy is pretty obvious. They're essentially gonna subsidize this, um, until they have as many engineers dependent on their, uh, tooling as possible with agents in their CI and background Claude agents and looping agents and all these things where it feels like every new feature and marketing push from these labs seem to be

[7:17](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=437s) a new workflow to standardize on, to use even more tokens and to be locked in even more. Um, and then inevitably the price gouging, uh, once they've got you trapped where your developers can't work without the tools.

[7:29](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=449s) And this isn't theoretical. We're seeing this happen live with some of the customers that we talk to. Um, so just a quick show of hands, how many of you basically stop working whenever there's a Claude outage

[7:44](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=464s) or a GPT outage? Yeah, same. [laughs] Um, I think that's a reason why we've seen Anthropic and OpenAI go from being API businesses to investing so much into the application layer, um, is because they know that that's where they can set these sorts of traps and build their moat for the day that these models inevitably become a commodity.

[8:06](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=486s) But I don't actually think the strategy is gonna work, um, and that's the message I wanted to get across today, uh, that this feels very shortsighted. And what we're noticing happen, um, in the world is that it doesn't matter how many features your CLI agent has, uh, developers and businesses will just jump to whatever offers them the

[8:24](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=504s) best value for their dollars. So if we look at current open-weights models, many of which are, are built in China, we'll notice that although they've lagged behind the American closed source competitors, we're at an inflection point where raw intelligence lead doesn't matter as much anymore, um, because these models are powerful enough where you don't always need the

[8:42](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=522s) best one for all your work, and that cost is becoming extremely important to these businesses that have kind of turned a blind eye until now.

[8:50](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=530s) And I think we all kind of feel it, that to get the best output from these models, it's more a problem of what context and tools you give the agent access to and less about its raw intelligence.

[9:00](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=540s) With the right AI native development infrastructure with project skills and rules, um, systems of verification and quality gates, um, even a mediocre model can produce similar results as a more intelligent model.

[9:14](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=554s) It just might take more tokens. Um, the intelligence is better placed in the system and guardrails around the model so that you don't have to be as reliant on the model or your end developers' responsible use of the model itself.

[9:28](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=568s) So we recently shared, uh, an anecdotal experience where we were skeptical of the benchmark saying that GLM was better than Opus. So we tested them on a real bug from the Cline repo, and while both models fixed the issue, GLM was the winner in terms of cost and code quality.

[9:45](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=585s) So GLM used twice as many tokens but only cost half as much. Opus finished faster. It used half as many tool calls. But GLM cleaned up dead code and verified that the build compiled before completing while Opus didn't.

[10:01](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=601s) It left a bunch of type errors, and it broke the production build. Um, and so that gave us the sense that GLM was trained to spend more tokens verifying its output, um, which is fine because the tokens are cheaper under- anyways, and it's really the end result that matters.

[10:18](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=618s) And because these open-weights models can deliver the same output, although at more tokens, we're seeing signs of the industry adopting and standardizing on these models. So this is Brian Armstrong, the CEO of C- of Coinbase, saying that they've, uh, defaulted to using GLM and Kimi in their internal LLM gateway and that this has cut their AI spend

[10:38](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=638s) by nearly half, uh, while their token usage continues to grow. And I think we'll see other businesses building their own internal tooling and routing to work with these agents in the most dollar efficient way for them, even if it means not having access to the latest new feature in something like Claude Code.

[10:57](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=657s) We're seeing the same thing that happened with Open Compute fifteen years ago. So, uh, just a quick history lesson for those that haven't heard of this. In 2011, Facebook was just getting started on building out, um, stuff like distributed computing infrastructure and data centers.

[11:13](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=673s) Um, but by the time they built it, Amazon and Google already beat them to the punch, so it wasn't a competitive advantage. So Mark said, "All right. Let's just open source it and see what happens."

[11:23](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=683s) Uh, so they took the designs for their data centers and their servers and their networking and cooling racks and all the physical hardware they spent all this energy and, and money, um, building, and they just gave it away.

[11:33](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=693s) They published, uh, schematics and CAD files and everything, uh, and called it the Open Compute Project, and what they saw was that the entire supply chain reorganized around it.

[11:44](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=704s) So before Open Compute, every company designed its own proprietary servers, so manufacturers were doing small production runs, um, of, like, very custom hardware, which was expensive. But when Facebook's designs became this sort of shared open standard, suddenly everyone was ordering the same thing, and manufacturers could do these massive standardized production runs commoditizing these components, um, so no

[12:07](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=727s) single vendor could charge a premium. And the price of everything came down, uh, for the whole industry, including Facebook itself. Um, and so what Facebook found was by giving these designs away, they created the market that drove their own costs down and saved them billions of dollars down the road.

[12:23](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=743s) And so I think the lesson, um, taught here is that the industry will adopt and standardize on something that they can build on top of, even if it isn't the best thing.

[12:34](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=754s) And with how much CapEx we've locked in for the next five years for AI infrastructure build-out, open-weights models are only going to get cheaper. Um, there's estimates that we'll spend nearly three trillion dollars and create, uh, over a hundred gigawatts of new data center capacity by 2030, um, roughly doubling, uh, global capacity today.

[12:54](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=774s) And hosting providers like Baseten and Fireworks, their whole purpose is to beat each other, uh, to beat the competition. So they'll use infrastructure efficiency gains like dedicated hardware and caching and batching volume tricks and inference-specialized silicon to drive costs down even more.

[13:10](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=790s) And, um, by 2030, the estimates are that inference on a one trillion parameter LLM will cost, um, ninety percent less than it does today.

[13:20](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=800s) We're seeing, uh, the same sort of cost-cutting tricks that commoditized the cloud ten years ago, um, happening in inference. Uh, so in 2014, uh, Google, uh, at their GCP Live March announcement said that they would cut compute by thirty-two percent and storage by sixty-eight percent.

[13:39](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=819s) Um, and AWS fired back with similar cuts within, within days, and this was AWS's forty-second, uh, price cut at that point. And so from 2015 onwards, once raw compute and storage were a commodity, the hyperscalers stopped competing on, on it as much and started competing on other things like databases and serverless.

[13:58](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=838s) And I think because open-weights allows these host providers to compete and cost, uh, optimize so aggressively, we'll see mass adoption of these foreign open-weights models, um, because when dollars are involved, the markets are extremely efficient, and the absurd API costs that these closed labs charge just won't be worth it anymore for most knowledge work.

[14:22](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=862s) Uh, and so this is me just, like, humbly requesting the American labs to take open-weights more seriously, uh, because before we know it, all this infrastructure that we're investing in could be built on foreign models that take the world by a storm and make GPT and Claude irrelevant.

[14:39](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=879s) Um, mindshare and adoption is incredibly important, and there's a chance that if the foreign models become the standard, there won't be a reason to switch back to GPT or Claude or Gemini, no matter what the marginal improvements are.

[14:51](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=891s) And then we lose control over the development of this technology. Um, and who knows where the world is, is headed if we don't have the likes of Anthropic and OpenAI to invest so heavily into safety research in ways that perhaps these other labs wouldn't.

[15:06](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=906s) I think the development of a technology this transformative is deeply, uh, tied to the ideals of the people and the nation that's building it. And to instill those values in the future of this technology, we need to keep the lead.

[15:19](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=919s) So I don't mean we need to open source our research. I think that's what gives us the lead. I think we need to open-- we need to open up our models and, uh, start releasing more open-weights models, which as we know, are not nearly as useful.

[15:33](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=933s) You can use, um, and extract the traces and, and train more, you know, your copycat models on them more easily, but not in a way that can leapfrog. And

[15:45](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=945s) this would make models, uh, more usable by the industry in a way that allows more competition and adoption and better price and value for customers, and I think that's how we keep our lead during this very critical moment.

[16:01](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=961s) Me and Cline believe so much in this open-weights future that we launched, uh, an open-weights subscription plan earlier this week that through volume, uh, based discounts and partnerships, uh, with inference host providers, we can offer significant discounts compared to paying for these models, um, at direct API costs.

[16:18](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=978s) Um, and we plan on continuing to increase the usage quota for these models as they become cheaper. Uh, you can sign up at that link, cline.bot/pass, uh, if you'd like to get a feel for how far models like GLM and DeepSeek have come and how you don't need the most expensive closed frontier model access to get work

[16:36](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=996s) done anymore. Uh, Cline is also, uh, open source, and you can bring your own API key and use any other provider. Um, you can use it on your CLI and VS Code and JetBrains.

[16:47](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=1007s) Um, and, uh, yeah, we continue to, um, add the newest models whenever they're released. So it's a good way to get a feel of, like, how much better the latest, newest model is, especially with open-weights because, um, you can't access those with like the Claude or ChatGPT subscriptions.

[17:07](https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=1027s) Uh, cool. That is my presentation. Thank you. [audience applauding] [upbeat music]
