Raven Protocol 🐦‍⬛

Raven is developing a distributed network of compute nodes for Artificial Intelligence and Machine Learning. A decentralized computing protocol built on #BNB

[Raven Protocol 🐦‍⬛ @raven_protocol]

🏆Raven wins the @mozillabuilders Award at the Fix-the-Internet showcase.

Recognition by the top open source organization in the world that Raven Protocol is a core infrastructure project.

We demonstrated best traction and commitment to fix the internet!
https://t.co/9CEqWOAiAB

[Raven Protocol 🐦‍⬛ @raven_protocol]

One of the biggest hidden costs in LLM fine-tuning is optimizer state memory.

New on arXiv: TACO uses a ternary optimizer to reduce that overhead without sacrificing convergence.

Worth a look for anyone working on fine-tuning.

https://t.co/SyZruJiQ4X

[Raven Protocol 🐦‍⬛ @raven_protocol]

@kevalgala03 gRAVEN!

[Raven Protocol 🐦‍⬛ @raven_protocol]

2/ Ravnest was built specifically for this environment.

Model parallelism across nodes. Gradient compression for consumer networks. Fault recovery that steps in before a node drop stalls the session.

The infrastructure layer for open source AI fine-tuning.

[Raven Protocol 🐦‍⬛ @raven_protocol]

1/ Open source developers fine-tuning large models on their own hardware hit the same three walls and most frameworks dont have answers for any of them.

- Memory limits across nodes
- Network bandwidth bottlenecks on internet
- Node failures mid-run that force costly restarts

[Raven Protocol 🐦‍⬛ @raven_protocol]

@NVIDIAAIInfra Idle GPUs are expensive. Better orchestration becomes even more important as AI workloads spread across different types of hardware.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@not_ellington Specialized chips need smarter orchestration.

Distributed compute will play a big role in making heterogeneous hardware work together efficiently.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@rohanpaul_ai A 500 MW build shows how quickly AI infrastructure is moving into gigawatt-scale planning.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@ShanuMathew93 The equipment lead times stand out.

Even with capital ready, power and hardware bottlenecks can delay deployment for months.

Distributed compute offers another way to access capacity while new facilities are being built.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@levie The coverage point is key.

Agent swarms could help security teams test far more code paths, but running them at scale also creates a real need for reliable, distributed compute capacity.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@sonnet_xu Persistent workloads make GPU ownership easier to justify.

Distributed compute adds another option for the capacity that sits idle between those workloads, helping owners put more of their hardware to use.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@TencentAI_News The jump after exploration is a useful signal.

It shows how much performance can depend on an agent’s ability to test hypotheses and learn from feedback, which makes the environment itself a key part of evaluation.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@NVIDIAAIInfra @OpenComputePrj Open standards make it easier for different parts of the AI stack to work together.

As infrastructure scales, that flexibility also matters for connecting diverse hardware across distributed compute networks.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@macropaperr Falling token prices make utilization a bigger part of the equation.

Distributed compute could help put underused hardware to work and give AI workloads more flexibility in how they source capacity.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@satyanadella Custom silicon at multi-gigawatt scale is a serious infrastructure commitment.

The tokens-per-watt gains will be especially important as inference demand keeps growing.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@ShanuMathew93 The gap between forecasted capacity and actual deployment is where distributed compute becomes relevant.

A more flexible hardware network could help absorb demand without every workload depending on new data center capacity coming online.

[Raven Protocol 🐦‍⬛ @raven_protocol]

@Jonsid The data-deployment loop also depends on having repeatable environments to test each improvement.

Distributed compute could help scale those environments and keep evaluation and training moving as workloads grow.

[Raven Protocol 🐦‍⬛ @raven_protocol]

3/ When you can't get more power, the only way to run more AI is to squeeze more out of what you have.

Power is the new bottleneck. Efficiency is the new benchmark.

For teams training on distributed consumer hardware, the power is already there.

[Raven Protocol 🐦‍⬛ @raven_protocol]

1/ The AI chip marketing playbook just changed, and it tells you exactly where the real constraint is.

-- In 2022: “9x faster training, 30x faster inference.” Chip vs chip. Power barely came up.

-- In 2026: every major launch leads with megawatts, not speed.

[Jessie Dong @jessiedong_]

power is everything now, and you can see it in how AI chips are being sold. we saw it again today with the launch of @Microsoft’s new AI chip, maia 300

a few years ago, companies marketed AI chips on how fast they were. now it’s also about power. microsoft’s new chip is marketed as getting 40% more AI output per megawatt than leading gpus

in 2022, nvidia launched the h100 as “9x faster training, 30x faster inference” than the last chip. the comparison was chip vs chip, and power barely came up

now every chip maker pitches in the framing of power:
- @nvidia: "vera rubin does up to 30x more per megawatt than its last generation"
- @awscloud: "trainium3 makes 5x more tokens per megawatt than trainium2"
- @Microsoft: "maia 300 makes 40% more per megawatt than leading gpus"

how should we think about old chips though? for data centers, is it worth swapping old ones out even if they're still usable? if you can’t get more power, the only way to run more AI is to get more out of the power you already have

[Raven Protocol 🐦‍⬛ @raven_protocol]

2/ Every major chip maker now pitches in power efficiency:

- NVIDIA Vera Rubin: 30x more per megawatt than last gen
- AWS Trainium3: 5x more tokens per megawatt
- Microsoft Maia 300: 40% more per megawatt than leading GPUs

[Raven Protocol 🐦‍⬛ @raven_protocol]

4/
- OpenAI plans $856B in compute and AI infrastructure spending through 2030
- Perplexity continues deploying GPT-6 Astra in production systems, including software modification and monitoring

[Raven Protocol 🐦‍⬛ @raven_protocol]

1/ Last Week in Open Source AI & Distributed Compute

- OpenAI's annualized revenue reportedly reaches $50B, around $20B below earlier reports
- OpenAI plans $856B in compute and AI infrastructure spending through 2030
- Microsoft unveils Maia 300, its next-generation AI chip

[Raven Protocol 🐦‍⬛ @raven_protocol]

3/
- OpenAI fires three safety researchers following an investigation into alleged policy violations involving sensitive information
- OpenAI moves to finalize contracts with third-party AI safety assessors