The service uses your university login (likely) and supports standard API connections. This system has already handled 11 billion tokens for hundreds of users. It bypasses the usual compute queues to provide fast results. I wrote a full guide on how to access these resources and get your team started.

Read the full article on Substack:

https://open.substack.com/pub/chatjpt/p/low-friction-and-no-cost-the-federated?r=12gd68&utm_medium=ios&shareImageVariant=overlay

For someone looking to build AI research agents on GCP, this is huge. It means semantic search across the world’s medical literature without the massive data engineering overhead (or the eye-watering compute bill) of maintaining your own embedding pipeline.

But as I dove into this “adventure” (which felt a bit like a fool’s errand at first), I realized that the “happy path” Google shows you in their documentation comes with a hidden tax.

The 115 GB Wake-Up Call

I followed the “Getting Started” guide, copied the sample SQL, and ran a query for one of my researchers. It worked beautifully. Then I looked at the execution details: 115 GB scanned for a single query.

At BigQuery’s on-demand rates, that’s about $0.70 per question. If you’re running a serious research project or powering an agent that loops through a thousand queries, your research credits will go up in smoke before you’ve even reached a conclusion. Leadership generally doesn’t get “optimistic” about AI when the bill arrives.

The “One-Line” Fix

The problem isn’t the vector search itself; it’s the SQL structure. Google’s example asks the database to perform vector search AND return the massive article_text column in the same swoop.

I tried a little experiment: I deleted one line of code — base.article_text — and ran it again.

Original Query: 115 GB

Lean Query: 13 GB

By fetching only the Article IDs, titles, and authors first, I dropped the cost by nearly 90%. You can always fetch the full text for the top 10 results in a second, tiny, targeted query. And you can probably do way more to shrink it…

FinOps Batteries Not Included

The lesson here is simple: Cloud providers are great at showing you how to get results in five seconds, but they don’t always show you the most cost-effective way to do it. When building AI agents and cloud infrastructure, you have to look past the “easy” code.

I’m still exploring how to plug this into MCP (Model Context Protocol) to see if I can turn this into a fully functioning researcher agent with A2A and MCP.

Read the full technical breakdown and see the code over on my Substack here:

Cool that Google embedded the entire PubMed dataset in BigQuery for semantic search, but....

DISCLAIMER: I am relatively new to GCP and BigQuery. If I’ve misinterpreted these scan results or if there’s an even more efficient way to query the PubMed dataset, please let me know in the comments! We’re all learning here.

Here’s the situation:

I’m a higher ed IT leader. Twenty years in cloud, data, security, governance. The usual. I’ve spent my career learning how to keep large organizations in higher ed running while the technology underneath keeps shifting. It’s fine. It’s good, actually. I think I am good at it.

But AI changed something for me.

Not in a vague “the future is coming” way. More like a specific dread. If I don’t get on top of full-stack enterprise AI myself, the actual building of it, I’m going to become one of those leaders who talks about technology but doesn’t really understand what’s happening anymore. I’ve met those people. They are irrelevant and become the butt of jokes. I don’t want to be that. AI will do just fine at that.

At the same time, I know I have gaps. Presence. Communication. Storytelling. The stuff I’ve been able to partially outrun so far. Not forever, though.

So I decided to work on all of it. At the same time.

Stage Academy, for presence and public speaking: https://stageacademy.mykajabi.com/

ContentCreator.com, for building a content habit and videography skill (and new hobby): https://www.contentcreator.com/

The AWS Generative AI Developer course https://aws.amazon.com/certification/certified-generative-ai-developer-professional/: I don’t really care about the cert. I want the structure, the outline, something to pair with actually building things and sharing what I learn. I know Langgraph, some ADK, and a good bit of Azure and AWS already. But just a little bit about a lot. I want to go DEEP and pair with my cloud experience.

(No affiliate links. No pitch. I’ll give honest reviews later when I have something useful to say.)

OH YEAH…there is also a goal to put myself out there more. To build a personal brand. Overcome my fear of sharing publicly. I need all these things to succeed there, right?

The question I keep circling is whether this is smart pressure, slow-motion self-destruction, or just accelerating procrastination.

Why I stacked all this on purpose

The easy explanation is that I got overexcited and signed up for too much. But that’s not the end of it.

Underneath it is a fear. Becoming obsolete and falling behind being in a sluggish and often lagging world of higher ed. Watching AI reshape everything and just kind of… waving at it from the sidelines while delegating the real work to people who actually get it or throw up my hands in a “oh silly higher ed” kind of way.

AI is moving fast enough that staying adjacent doesn’t work. Reading about it, vibe coding, trying to stay up to date, hiring smart people. Oh yeah and the pace. It’s hard not to BURN OUT on that alone, right?

But also going into this new world in technical leadership and being successful feels like you HAVE to double down on the interpersonal. You can’t just be the technical person. You have to walk into rooms with faculty and CFOs and skeptical board members and actually move them. Build trust. Explain things in plain language. That requires presence, clarity, the human stuff.

So I ended up with this picture in my head: technical depth, enterprise context, human connection. And I didn’t want to improve one piece at a time. I wanted to pull on all of them at once and see what happened. And share about it.

That’s the optimistic framing, I guess.

The tax of always upgrading

When every spare hour goes toward becoming future-proof, not much is left for just existing.

I’ve noticed I have less patience for hobbies that don’t connect to my goals. If I can’t tie something back to leadership or AI or videography, it slides to the bottom.

Even rest starts to feel tactical. I’m resting so I can perform better tomorrow. Relaxing becomes another input to the productivity machine.

Weird thing: the more I learn, the more behind I feel. Every course opens five more rabbit holes. Every project suggests three more things I should probably understand and build.

SIGH.

Why I’m putting this out there

I suspect a lot of people in tech or higher ed or leadership generally are doing something similar. Stacking building and learning and side projects on demanding jobs, calling it growth, privately wondering if they’re running scared. Feeling FOMO at those more productive and using AI better, or speaking on camera better, or presenting better.

Others just feel alive right now riding the wave.

I’ve always wanted to surf.

If you’re in that spot, the only question that cuts through my own noise is this: if I stopped adding new things for six months and just executed on what I already know, would my future actually be in danger?

For me the answer is no.

Which means this isn’t about survival. It’s about identity. Who I’m trying to become.

I don’t know if that makes it better or worse.

I’ll share reviews of these programs once I’ve given them a fair run. For now I’m just in it, watching myself, trying to notice where the line is between stretching and hollowing out.

Haven’t found it yet.