Chief of Tokens · Senior Director · AMD · University of Waterloo

MARIO
FILIPAS

The goal isn't AI everywhere.
It's AI where it counts. Value per token, not cost per token.
Aim the tokens where the returns are.

I lead 150 engineers across Canada, Serbia & China, building GPU virtualization software for AMD's Instinct AI accelerators. I move with urgency, and I run on AI. We're powering the next generation of AI - and defining how to get the most out of it.

Read the manifesto Get in touch →
150+ Engineers Led
3 Countries
3 Hypervisor Platforms
Upside if you move now
Manifesto
AI will not be scarce.
Judgment will be.
Intelligence is becoming abundant. Everywhere. Nearly free.

The advantage won't come from having it.
It will come from knowing where to aim it.

AI makes everything faster.
But faster in the wrong direction is just expensive motion.

The future belongs to those who know what matters - and point the machines at it.
Build judgment.
Aim intelligence.
Move the business.
Act accordingly…
- Mario Filipas, Senior Director AMD · GTA, Canada
About
Disruptive tech.
Built by teams
defining
what will be.

I'm a Senior Director at AMD, leading cloud GPU software development for the Instinct and MI-series AI accelerator line. My teams own the driver stack that enables GPU virtualization across KVM, Hyper-V, and VMware - the software layer that lets hyperscalers and enterprises partition and run the world's most powerful AI hardware reliably at cloud scale.

I lead 150 engineers across Canada, Serbia, and China - and I'm pressing them to adopt AI ahead of the industry, out at the edge where the playbook hasn't been written yet. Test infrastructure rebuilt for the age of AI-generated software. Bug turnaround falling and feature throughput climbing, year over year. My job is to build the environment, direction, culture, and people that make that pace sustainable.

I graduated from the University of Waterloo in Computer Engineering. I think entrepreneurially. I believe the gap between where AI is today and where it's going in five years is the biggest opportunity of our professional lives - and I'm not willing to watch from the sideline.

Role
Senior Director, Cloud GPU Software Development
Company
AMD Instinct & MI-series · Cloud & Virtualization
Team
150+ engineers · Canada · Serbia · China
Domain
GPU VirtualizationKVMHyper-VVMwareCloud Drivers
Products
MI355XMI350XMI400 SeriesHeliosROCm
Education
BASc Computer Engineering · Minor in Management Science · University of Waterloo
Personal AI
Claude CodeCursor Agentic coding · terminal-first · daily driver
Mindset
EntrepreneurialAI-NativeSystems ThinkerBuilder
The scale
150+
Engineers

Three countries, thirteen time zones, one direction. Leading at this scale means building the culture, the process, and the technical bar simultaneously - and making sure AI is embedded in how every team works.

3
Hypervisor Platforms

KVM, Hyper-V, VMware. Each platform is a different world. Shipping reliable, performant GPU virtualization support across all three - for the most in-demand AI silicon on earth - requires zero margin for error.

What's at Stake

GPU virtualization sits at the exact intersection of AI infrastructure, cloud economics, and semiconductor dominance. The next decade of computing is being built on this stack. Right now.

Experience
Present AMD Greater Toronto Area
Senior Director, Cloud GPU Software

Accountable for the driver stack that lets cloud providers deploy and partition Instinct MI-series GPUs at scale - the foundational software layer for how the world consumes AI compute. It's also the proving ground for everything I write about: an engineering organization going AI-native in production, not in theory.

AMD Instinct Engineering Leadership GPU Virtualization Distributed Org Design AI-Native Culture
Prior AMD Canada
Sr. Manager → Director

Led the Windows Software Platform Group: the kernel-mode driver and the full software stack delivered to Microsoft for Windows operating systems running AMD GPUs. Built and scaled the teams, drove platform strategy, and developed the operational muscle to run distributed engineering organizations across time zones, cultures, and technical domains.

Windows Platform Kernel-Mode Driver Team Building Platform Strategy Program Management
Foundation University of Waterloo Waterloo, ON
BASc Computer Engineering · Management Science Minor

One of the most rigorous engineering programs in North America, paired with a co-op model that means graduating with years of real industry experience. Computer Engineering built the technical depth to understand what teams are building. The Management Science minor added the second lens early: not just how to build, but which problems are worth the resources.

Computer Architecture Operating Systems Embedded Systems Management Science Minor Co-op Program
Thinking
// 016 · Satire
A Press Release From a Company Doing AI Exactly Wrong

Apex Dynamics gives up on "solving with AI" by hiring a consultant to tell them what they should already know.

Read →
// 015
Value per Token, Not Cost per Token

Tokens aren't a cost problem. They're a guidance problem. The winners won't have the lowest AI bill - they'll generate the most value per token.

Read →
// 013
AI Where It Counts

Not seeing the gains and spending too much on tokens are the same problem. Stop applying AI everywhere. Aim the tokens where the value is.

Read →
// 001
Teams That Use AI Are Outperforming Teams That Don't

This was a prediction. Now it's happening. What it means for engineers, teams, and companies who want to still be relevant next year. A call to action.

Read →
// 012
Your AI Strategy Has a Blind Spot at the Top

Pilots everywhere, gains nowhere. ICs accelerate tasks - only senior leaders can redesign how work flows across teams. You can't redesign what you don't understand.

Read →
Read all 16 essays →
How I work · Now

I work in the terminal.
With Claude Code.

I got here by iterating. OpenClaw - an open-source autonomous agent powered by Claude - ran my calendar, my inbox, and my daily workflow for months. It proved what agentic AI could do. It also cost me my evenings: endless tinkering just to keep it working. Claude Code flipped that ratio - the same agentic power, in the terminal, and it just gets things done. This site included. Cursor is my daily driver at work, and now I'm running it at home too.

The tools will keep changing. The principle won't: aim the tokens where the value is. AI-native doesn't mean AI everywhere - it means AI integrated into how the work actually flows, where it moves the outcomes that matter.

claude code · mario@filipas - now
/status
# Jul 2026 · last ship: this site · Claude Code
shippedfilipas.com + 16 essays + RSS
livebitcoin.filipas.com · power-law tracker
day jobInstinct virt · 150 eng · 3 countries
stackClaude Code · Cursor · terminal-first
thesisvalue per token, not cost per token
/recent
016  Apex Dynamics satire press release
015  Value per Token, Not Cost per Token
013  AI Where It Counts
Beyond the work
mario@filipas - beyond the work
ls ~/beyond-the-work
bitcoin/Bitcoin Power Law Tracker - the model from essay // 005, live
christianity/Biblical Genome - a networked map of biblical reality
history/The Filipas of Cres - eleven generations on one Adriatic island
# all built the way I work: AI-native, terminal-first

LET'S
TALK.

If you're building at the infrastructure layer, thinking seriously about AI adoption, or working on something that matters - I want to hear about it. Based in the Greater Toronto Area.