founder @swiftinference @tizeti YC W17 ex @microsoft ex @pwc ex @exxonmobil
California, USA
Joined September 2009
- Tweets2.1K
- Following543
- Followers2.5K
- Likes3.2K
Pinned Tweet
It’s my birthday. Instead of gifts, I want data. I just shared something I’ve been building since the start of this year. The full story is here. And I need your help proving there’s a Nigerian in every country on earth.
Ken retweeted
Big news for our developers! We are shipping a developer preview that dramatically expands the capabilities of all apps to integrate into Siri. So excited to see what you folks do with this.
developer.apple.com/news/?id…
Ken retweeted
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
Ken retweeted
We’ve raised $1.3M from YC & Afore Capital so you never have to type or speak again.
Typing is slow. Talking into your computer is awkward. Subvocal fixes both
Our glasses give you the speed of speech without you making a sound or touching a keyboard.
It tracks your internal speech muscles with non-contact sensing and decodes silent, almost invisible mouthing into text at more than 200 words/min.
Pre-orders are now live at:
subvocal.co/preorder?utm_sou…
Orchestrating, routing and scheduling etc are all important in inference and is usually the difference between a high and low ROI on your compute.
In a 128-rack NVL72 cluster, a node goes down about every 5 hours.
That’s not a bad cluster. It’s the interruption rate Meta reports for its Prometheus cluster, spread across 2,304 nodes. Each rack fails roughly every 27 days. Here’s why that matters.
A rack is 72 GPUs on one NVLink fabric, so a job that stays inside the rack runs fast, while the same job spread across racks runs up to 43% slower. The obvious move is to fill every rack with an 18-node job. Then a node goes down, and the whole job stops. With 2% of nodes down at any time (what we see in production), 30% of racks have a node down and can’t even start one.
Now try something that sounds wasteful: give each job 16 nodes and leave 2 idle as spares. When a node goes down, a spare in the same rack takes its place and the job keeps going. A rack only drops out when 3 nodes are down at once, which is 0.5% of racks.
Across 128 racks, that’s 2,037 usable nodes instead of 1,602. Leaving 256 nodes idle bought 27% more capacity.
We wrote a blog reporting our experiences with NVL72 in production . We built the SkyPilot Topology-Aware Scheduler (TAS) to smartly and automatically place jobs optimally and recover them after failures. It keeps jobs rack-local, binpacks, swaps failed nodes for in-rack spares, and frees entire racks by moving fragmented jobs out of the way.
In large training runs, SkyPilot TAS can raise training throughput by >14% ($400K+/day savings) and usable capacity by 27%.
1
160
Ken retweeted
First of its kind gas turbine that generates 40% more power and creates its own water for cooling
Today we launch @VirajAero (YC F26).
We built a turbine to fly. Now it powers data centers: it makes its own cooling water from the exhaust, gets 40% more power from the same fuel, and gets permitted in weeks.
No water connection, no wait for the grid.
Ken retweeted
In a 128-rack NVL72 cluster, a node goes down about every 5 hours.
That’s not a bad cluster. It’s the interruption rate Meta reports for its Prometheus cluster, spread across 2,304 nodes. Each rack fails roughly every 27 days. Here’s why that matters.
A rack is 72 GPUs on one NVLink fabric, so a job that stays inside the rack runs fast, while the same job spread across racks runs up to 43% slower. The obvious move is to fill every rack with an 18-node job. Then a node goes down, and the whole job stops. With 2% of nodes down at any time (what we see in production), 30% of racks have a node down and can’t even start one.
Now try something that sounds wasteful: give each job 16 nodes and leave 2 idle as spares. When a node goes down, a spare in the same rack takes its place and the job keeps going. A rack only drops out when 3 nodes are down at once, which is 0.5% of racks.
Across 128 racks, that’s 2,037 usable nodes instead of 1,602. Leaving 256 nodes idle bought 27% more capacity.
We wrote a blog reporting our experiences with NVL72 in production . We built the SkyPilot Topology-Aware Scheduler (TAS) to smartly and automatically place jobs optimally and recover them after failures. It keeps jobs rack-local, binpacks, swaps failed nodes for in-rack spares, and frees entire racks by moving fragmented jobs out of the way.
In large training runs, SkyPilot TAS can raise training throughput by >14% ($400K+/day savings) and usable capacity by 27%.
I just met @DeepgramAI CEO @deepgramscott and CTO Adam at #DeepgramSpeak exited about all the progress with Voice AI being showcased.
6
255
This is so cool
So one of our ambassadors @divagr1925 has been working on something cool
He found a way to use Qwen's Gated DeltaNet state as memory, and built a 32-bit RISC-V CPU directly into the model's weights. So instead of relying on predicting the answer to a calculation, the model can execute instructions internally. It can also run C and MicroPython programs.
He calls it Qwenputer, love the name btw. The results so far are pretty impressive. On one arithmetic benchmark, accuracy jumped from 48.6% to 100%.
Pretty cool to see developers experimenting with Qwen's architecture. Great work, my friend!
96
Last week, @SwiftInference received a ton of GPU compute inquiries, within 24 hours of announcing our Santa Clara node.
It reinforced our thinking: AI doesn't just have a GPU shortage. It has an infrastructure architecture problem. We spent the week working on the request and wanted to share my thought around stranded GPUs.
Here's what the industry is missing. Everyone is building data centers needing gigawatts of power, enormous capital and years to construct and then wondering why power is a bottleneck. What if we stopped putting all the world's AI compute in a handful of locations? Existing facilities can accommodate compute today.
At SwiftInference, we distribute GPU compute across locations and connect them using our intelligent routing software. We make running inference for companies feel like one unified platform, while behind the scenes, workloads are routed across distributed infrastructure based on capacity, geography, and latency.
It's simple, spread the power requirements across multiple facilities, reduce dependence on any location, and serve inference closer to users. Perhaps this seems straightforward given my telecom and offshore oil and gas background. During COVID lockdowns, we moved 5G and Wi-Fi equipment across borders. Think about that period, factories were shut down, logistics were frozen and governments prioritized ventilators and PPE. and yet we still got equipment operational.
Next, came the global semiconductor shortage and today, those constraints have resurfaced in the GPU market. Supply chain firefighting taught me something: you don't solve bottlenecks by waiting for perfect conditions, you design around them.
A VC suggested we start with a low-hanging fruit and warned it could take up to two years. We started with the hard part, and six (6) months later, we're live in Santa Clara with carrier integration and production infrastructure. Sometimes the impossible timeline reflects on who's doing the building.
As if the week wasn't eventful enough, there was the PHP2Shell cybersecurity crisis. Centralizing enormous amounts of compute and sensitive data creates concentration risk. Distributed infrastructure, when properly secured, can reduce the blast radius of an incident. This also extends to scope of outages too.
So here's my message to anyone with GPUs in a warehouse, waiting on power, colocation, networking, or deployment: Send them our way. Our Bring Your Own GPU service brings you online in days, not years. We'll handle the infrastructure, deployment, and you look like a rockstar to your management. We've been building and operating infrastructure for 15 years. Now we're bringing that experience to AI.
The GPUs exist, the demand exists, and the infrastructure doesn't all need to be in one place. Let's get them connected.
1
2
209
It's inevitable that new tech will change the paradigm and independent thinking from first principles for aspect of it from infra to model.
Yes..great researcher doesn't great entrepreneur
People ask: “Are you dunking on neolabs?”
Most neolabs will fail. That’s normal. But I believe more AI breakthroughs are coming.
I’m skeptical of those following the same playbook as OpenAI and Anthropic. Raising $1B and spending 50% on GPUs doesn’t make you a $10B company.
I’m rooting for the contrarians. Challenge scaling laws. Challenge transformers. Challenge how models learn. Challenge the next-token prediction paradigm.
My bigger concern is that brilliant researchers don’t necessarily make great founders. In my experience, researchers often struggle to turn their research into great products / sustainable businesses.
120
They should send it to us @swiftinference we will manage the compute for them and turn it to open ai compatible endpoints in 2 weeks for them swiftinference.ai/docs/byog Our bring-your- own-gpu service has collocation locked in across the US
2
173
Another Reason why you need @swiftinferenceto speed up inference.
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
94