Parker Rex
All videos
Parker Rex

I've Used Both Google and OpenAI's Full Stack - Here's Why One Is About to CRUSH the Other

May 19, 2025

Watch on YouTube@parkerrex

In this video, Parker weighs the OpenAI and Google full-stack bets, arguing that Google’s end-to-end stack and massive AI infra give it a decisive edge as the AI tooling landscape matures.

Key takeaways#

  • The “stripification” trend: AI toolchains are consolidating toward single platforms that cover data, models, and deployment needs—similar to how Stripe unified payments.
  • Google’s edge: Google already has a complete stack (infrastructure, data, models, and tooling) plus massive AI spend, TPUs, and an unparalleled data backbone.
  • OpenAI’s push: OpenAI is aiming to be the central hub for AI apps and agents, but Parker thinks they’re underestimating Google’s breadth and scale.
  • Agent race: 2026 is pitched as the year of the agent; 2027 could see AI pervading the physical world. Expect consumer-level guardrails and traceability to matter a lot.
  • Practical stance: For developers, Google’s Vertex AI, Gemini 25, Veo 2, and vector search offer a powerful, integrated path. OpenAI remains valuable for experimentation and specific tools, but not the default for full-stack development.

OpenAI vs Google: the landscape#

  • OpenAI strategy
  • Emphasizes agents and an “AI subscription” approach.
  • Aiming to be the home for AI apps but faces gaps in consumer-facing tooling and end-to-end guardrails.
  • Poised to push a broad, multi-tool ecosystem, but execution details (like UI and trace visibility) are still evolving.
  • Google strategy
  • Boasts the entire stack: data, model infra, tooling, and deployment, plus heavy investment in AI infrastructure.
  • VP/Vertex AI ecosystem includes vector search, data ingestion, and enterprise-grade tooling.
  • Strong productization around developer experience and traceability (with ongoing improvements anticipated).

Google’s stack and what to lean into#

  • Vertex AI and ecosystem
  • End-to-end tooling for model management, training, deployment, and orchestration.
  • Vector search capabilities for ingestion of PDFs, websites, and other data sources.
  • Gemini 25 and Veo 2
  • Gemini 25 as the flagship multi-modal model; Veo 2 as the next-gen image model.
  • In Parker’s view, Gemini 25 and Veo 2 outperform comparable OpenAI offerings in practice.
  • Storage and data foundations
  • Infinite or near-infinite storage concepts via scalable buckets.
  • Easy data ingestion pipelines for training and retrieval across enterprise data.
  • Developer experience
  • Web UI and code-first workflows exist, but there’s a learning curve for building agent-like workflows.
  • Guardrails, tracing, and visibility into agent decisions are identified as areas Google is actively evolving.

Agent-based workflows: considerations for 2026–2027#

  • Agent vs multi-step autonomous systems
  • Agentic workflows resemble a chain of steps with checkpoints and visibility into decisions.
  • Autonomous agents can drift in quality with multi-turn interactions; guardrails and traceability are crucial.
  • The need for guardrails and traces
  • Expect improvements in UI to show decision traces, decision points, and error handling.
  • A robust SDLC-like approach (with HL/HITL where needed) will be essential for production agents.
  • Practical implication for builders
  • Start with clear, auditable agent flows and plan for monitoring, logs, and guardrails.
  • Consider SDLC tooling like CLI-based pipelines to stitch together agent steps.

Practical takeaways for builders#

  • If you’re building AI apps, consider the Google stack first
  • Start with Vertex AI for model management and deployment.
  • Use Gemini 25 for multi-modal capabilities and Veo 2 for image tasks.
  • Leverage vector search and structured data ingestion to build robust knowledge bases.
  • Data ingestion and storage
  • Ingest PDFs, websites, and other data sources into a managed data lake (buckets) to fuel retrieval-augmented workflows.
  • Guardrails and visibility
  • Implement traces and step-by-step visibility in agent workflows.
  • Plan for error handling, rollback points, and human-in-the-loop (HITL) where critical.
  • Practical tooling tips
  • Python remains a strong glue language; leverage Google’s Python libraries and Vertex AI client tooling.
  • The OpenAI API is still useful for experimentation, but for full-stack development, the Google stack offers deeper integration.

Personal stance and recommendations#

  • Parker’s position: He’s all in on the Google stack for development and deployment, with continued but selective use of OpenAI for experimentation.
  • Why Google wins on the current trajectory
  • The breadth of the stack, the scale of infrastructure investment, and the data/compute advantages give Google a durable moat.
  • The ecosystem’s maturity (storage, vector search, model deployment, and traceability) creates a superior developer experience for end-to-end AI apps.

Quick how-to starter (conceptual)#

  • Example starting point with Vertex AI (high level)
  • Initialize your environment and project settings
  • Upload and index data (PDFs, docs, websites)
  • Build a vector store for retrieval-augmented workflows
  • Deploy a Gemini 25-based model and iterate with Veo 2 for images
  • Set up traces and guardrails in the agent workflow

Code snippet (conceptual starter)

# Minimal Vertex AI setup (conceptual)
from google.cloud import aiplatform

aiplatform.init(project="YOUR_PROJECT", location="us-central1")

# Pseudo-steps:
# 1) Upload data to Cloud Storage
# 2) Create a Vertex AI index / vector store
# 3) Train or deploy Gemini 25 as needed

What to watch for next#

  • AI infra bets: Expect more announcements around agent tooling, guardrails, and traceability from Google.
  • Adoption patterns: Watch how enterprises adopt “stripified” AI stacks versus best-of-breed component approaches.
Transcript

The OpenAI guys are saying a lot of stuff about agents, about where we're going, how 2027 is the year when AI takes over the physical world, how 2026 is the year of the agent. We're going to be talking about both the chair of OpenAI and then Santa the boy. So, I agree with most of the stuff and disagree with some of the stuff. So, in this talk, I'm not even going to let him talk. I'm just going to talk about the talk and it'll be a short video, but you should just go watch it and make your own opinions. If you don't the TLDDR of this one is they are stripifying the AI stack and I talked to a founder seedstage AI founder in San Francisco and it was funny to hear because they didn't recognize that we were in the same space where it's 2015 payments and I'm old so I did payments then and we were between using a bunch of different payment styles so it was like oh we have a for some of the restaurants oh we credit card for some of the restaurants. We have card on file which was like slightly different. So it could be like at the end of the month they swipe or every week they swipe. We had all these different terms. We had cash payments from customers. We had our dev team wanted to build Bitcoin. We didn't do that. But the point is it was very fractured. And then Stripe came along. They dropped payments. We went from using Brainree because Brainree merged with Stripe and we went from them to Stripe. Stripe was then just payments. But then we still had to figure out all the stuff around it. And so payments were still really fractured, meaning if you wanted to run reports, they didn't have Sigma, which is their reporting stuff. They didn't have radar for fraudulent stuff. So you had to go pick up another tool and figure that part out. You had to do all these things in order to like cobble it together. And then that left side of the Stripe UI started getting built out. So what started as just dev- payments from the Collison brothers quickly became the whole ecosystem. So, OpenAI is really pushing on that button, which is like they want to be like the home for all of it. I think that they're missing the boat a little bit in terms of how powerful Google is. Google has the whole stack. They don't have consumer side. They're pushing on that. But I'm going all in on Python and the Google stack because I've just used it. I've used the OpenAI SDK since the day that it came out. like that was a little while ago, but they've made a lot of progress. I just think that models are getting commoditized that they don't recognize, like I said, that Google just has everything. And as a builder, knowing that Google's putting 70 or 90 billion of spend into AI infrastructure is just something that nobody else can do. Like you don't see the CEO or the board of Alphabet going to Saudi to cut deals and raise money. Elon doing that, Sam doing that. Of course, Jensen from Nvidia's there because he's the king of the chips. Google doesn't have to go because they have an ad business that is absolutely smashing and they can then fund all the infrastructure which people don't think about for deploying. Like you can go in and they don't have the web UI set up for this. You have to actually learn how to code and code the stuff for agents. that their agent setup is amazing and you can do a whole lot of stuff. It's very easy to understand and beyond that just the vast amount of things of options that are in there. Like I should totally be a Google rep. Shout out Google because it is fantastic. And let's just look at it. So we have buckets. This is infinite storage. You can mount one of these to your computer if you would so please, but it'd be like $23,000 a month if you actually filled up the full pabyte or two pabytes. But you can do selfcleing stuff. So I treat this as like temp storage once videos are processed. But you have buckets and then you have all these other things. So this is the stripification of AI that's happening right here. This one's going to win. It's just going to happen. And why? Because Google has the TPUs, they have the science, they have the money, and they also just have the internet. Like they've trained so much information since 2005 that it's just bananas. And when you want to go and do something, you jump into Vert.ex AI and it's okay, cool. Like, here's everything from OpenAI, but more. I think the only bit that's missing from this right now, which they're actively working on, they might announce at IO this week, is the pieces around like guardrails and like a web UI where you can see the traces. So if I launched an agent, and I hate to say it, I'm just going to use it because it's easy, but the travel agent because everybody uses it as the example. You can see like where decisions were made in an agent flow to for clarity. An agentic workflow means that you have let's call it seven steps or seven nodes in a row. You can go back to another video on building like AISDLC which is a software development life cycle which is a CLI that like chains together all those things. Those steps in both the software development life cycle or being a travel agent can be agentic meaning that each step has a stop gap for you to see what's going on. But then like full agent is it's autonomous because you have the information and you're giving it the ability to go and do its thing. There's still issues with multi-turn. So the quality of it decisions goes down drastically if you have like multi-turn going on or if there's a lot of revisions. But I think as soon as Google adds this then it's game over and they just are cranking. They have Veo 2 the new image model that's coming out. The existing image or video model is already better than OpenAIs. Gemini 25 is flying. If anyone doesn't know how to use it, learn how to use it. I think the only pe people that think it's not good are those that are like cursor stands that don't understand that cursor has not figured out how to work with it. But things like that, things like vertex AI search, vector search where I can take, oh, I want to learn more about software development life cycle. So, I'm going to just feed it all these PDFs, feed it all these websites. PDF isn't that fancy. That's just rag. But I can feed it a whole website and be like, yo, I want to take every Wikipedia page that mentions the software development life cycle and throw it on there. It's bananas. So, what Sam goes into in this talk is more about like the future of their striification strategy, going into how they want to be the core AI subscription. He's like very vague. He's been trained from his PR agency about like whenever you're asked a question about something that could be proprietary, just answer with better models. That's great. He also says that next year is going to be the year of the agent. On the other side of the coin, we have Brett. So Brett is starting his or started his own company and the thesis is that every company is going to have one agent. Forget the website. It used to be about websites, now it's one agent. It's funny because he's basically he's a really good salesman. He's a really good marketer and he's a really good engineer which is rare. He was a CTO of Facebook at like age 27 and so he knows how to do all these things. He's multidisciplinary which is a whole video in itself. I think that being multid-disciplinary is required now to be 10x because you can and you can learn anything. But Brett goes into the fact that you need to have an agent. Most of the things that they're selling is customer service. So if you go like every single customer success story on there whether it's Sonos or I don't know Clavio will be about how their seesat or customer service score essentially went up by using a product like Sierra. But the gist of this talk is basically yeah agent this is the year it's going to be awesome for B2B. It does make sense like his product's probably awesome. I'm not saying it doesn't work, but it's also interesting to see how he can blend the marketing with the technical because on the marketing side, like no one wants to hear multi-agentic workflow with HITL or all that stuff. It's just like super jargony. So, he's able to portray it and display it and speak to it like it's oh, it's just this one agent and then that makes it a lot easier for their sales. So, I in total like in in summary just want to encourage you to go watch these videos and come up with your own opinions, but then also be open to trying different tools. I still use OpenAI. I just don't use any of this stuff for development. And then I use like their obviously their web browser interface is awesome. They just added in the ability to run Python in it, but it doesn't do package management very well. But that's besides the point. Go check out the talks. If you learned anything in this video, make sure you subscribe to the channel, like the video cuz it's like the thumb thing. I don't know if you've ever seen it, but you should try it out. When you click it, it looks crazy. And then if you want to join our community, it's awesome. We've got people in there, I don't know, maybe a thousand times smarter than me. So, I just I'm the guy that yaps on the internet. And that's it for today. Hey, I'll see you in the next