Technically Monthly (Issue 1, April 2025)
Notes on Microsoft's quantum breakthrough, why AI cloud costs keep going down, how to build full-stack apps with AI, and more.
Hello loyal, handsome Technically readers. This is the first Technically Monthly, and it’s a big one. We’ve got:
4 brand, spanking new posts on Technically 2.0
A guest breakdown of Microsoft’s recent quantum computing announcement
A guest analysis of why AI cloud costs keep going down
A new term in the Technically universe: context window
Future issues will likely include a few other things, like bringing back the mailbag (should we?) and interesting threads to check out. But this issue does not include those things.
New posts on Technically 2.0
How to build full-stack apps with AI [free]
In the AI, it’s not that complicated knowledge base
A walkthrough of v0 by Vercel, a platform you can use to build full-stack apps without needing to write any code (in Vercel’s Next.js or any framework).
Thanks to Vercel for sponsoring the “AI, it’s not that complicated” series through the rest of the year, and helping us bring you some good stuff there for free. While keeping the roofs over our heads.
Breakdown: the observability tooling market [paid]
In the Analyzing software companies knowledge base
An overview of observability and monitoring companies: what products like Datadog, Splunk, New Relic and Elastic do, how they make money, and who they compete with.
Observability is a wild category with some 20+ legitimate options across open and closed source. How do they stack up against one another, and why do developers choose any particular one?
The category breakdown is a new format for Technically, specifically built for financial professionals trying to understand a broader category in the markets. We’ll be doing similar writeups for databases, AI, analytics, etc down the line.
What’s HTML? [paid]
In the Building software products knowledge base
Ever wish you could magically get any webpage to say exactly what you wanted?
Well now you can (sort of). This foundational post explains what HTML is, how it fits into the broader web ecosystem, and how you can use it to prank your friends (and enemies).
The not-engineers guide to building Technically [free]
A walkthrough of how we built the new version of technically.dev, and all the technical concepts (from webhooks and integrations to frontends and backends) that we needed to make the move from Substack to our own web app.
If you’re curious, we went a little deeper in this Substack note about how we used AI to organize the years-long Technically library of posts into cohesive knowledge bases.
Notes from the field: major technical news
Quantum bits & pieces: Microsoft’s Majorana breakthrough (?)
From Technically’s quantum computing correspondent, Conor Deegan of Project Eleven, an applied quantum computing and cryptography lab.
A few weeks back I covered What Quantum Computing Is. The TL;DR:
It’s an experimental technology that uses the weirdness of physics to do a whole bunch of tasks much faster than traditional computers. And it has the potential to solve problems classical computers find nearly impossible.
The problem is making it actually do stuff we want is hard (requiring temperatures colder than space, the control of tiny powerful lasers, and even then results are unpredictable).
Last month, Microsoft announced its Majorana 1 chip using "topological qubits," which they claim are more stable and scalable than existing designs. Sounds groundbreaking—but what does it actually mean?
What makes Microsoft's approach different?
Microsoft's approach revolves around Majorana fermions—particles physicists have hunted for decades, whose unique (and still hypothetical) properties make them resistant to environmental interference.
Current quantum computers struggle with stability because quantum information degrades easily. Topological qubits spread their quantum information across multiple particles in a clever way, making them more error-resistant.
Think of it like storing your passwords in several secure locations instead of just one. If someone breaks into one spot, your information remains safe elsewhere.
Skepticism remains
We've been down this road before. Microsoft claimed to detect Majorana particles in 2018, only to retract when other scientists couldn't reproduce their results. This history makes researchers wary, adopting a "trust but verify" stance on Microsoft's latest claims.
But proving topological qubits even exist isn't straightforward. Why?
They rely on subtle behaviors that only appear under ultra-low temperatures and carefully controlled environments.
Confirming these qubits is like solving a 10,000 piece puzzle—there are many false trails and misleading clues, so researchers have to rule out every other possible explanation first.
What to watch for
Even if Microsoft proves topological qubits exist, scaling into a fully functional quantum computer is going to be challenging. It’s one thing to create a handful of qubits in a lab, it’s another to reliably produce hundreds or thousands.
The next several months will be crucial. I'll be watching for:
Peer-reviewed research independently verifying Microsoft’s claims
Reproducible evidence showing real topological behavior + reduced error rates
If Microsoft’s claims are verified, it’d confirm a genuine breakthrough for quantum computing. The race is on, and the outcomes could reshape technology and society in ways we're only beginning to imagine.
AI Cloudonomics: Why AI is getting cheaper to run
From Technically’s cloud correspondent, Kenny Ning from Modal, a cloud compute platform that companies like Substack use to run their AI products.
OpenAI's API prices are in freefall (the good kind). From 2023 to 2024, the cost of processing a million tokens through their API has dropped from $36 to $0.25.
The same thing is happening with the powerful chips, called GPUs, that AI is run on. The hourly rate to rent an H100 GPU (the Ferrari of AI chips) on Modal has dropped 48% year-over-year, from $7.65/hour to just $3.95/hour.
Why is this happening?
Remember the cloud?
Cloud lets companies rent the servers (giant computers) they run their software on instead of buying them.
It's like using Uber instead of buying a car – you pay for what you use, when you use it.
The cloud compute market connects:
Compute-hungry companies (e.g. software-as-a-service companies, AI developers)
Cloud platforms that own the actual computers (AWS, Google Cloud, etc.)
It’s a huge market that is expected to reach $2 trillion by 2030, driven mostly by me asking ChatGPT to explain async await again ✨AI✨.
What moves the AI compute market
Here's what's driving these price drops:
More supply from surplus GPUs bought in 2023: Early in the AI hype cycle, startups and investors purchased large reservations of GPUs expecting a shortage. This turned out not to be the case, and many of those early startups have likely gone bust, so a lot of those early reserved GPUs are just now being released back into the supply.
Demand shift from training to inference: Demand for AI is growing, but we’ve noticed a shift in compute demand away from model training towards running inference (generating outputs from pre-trained models). Inference generally has less strict compute requirements compared to training, so why lease a Lamborghini when a Camry gets the job done?
More flexible ways to buy compute: You used to need long-term contracts with a cloud platform like AWS to get GPU access. Now you can:
Pay for unused capacity (that can be interrupted at any time 🤯)
What this means for you
When cloud prices drop, companies building AI tooling pass those savings back to you. Oh wait, just kidding.
They will actually likely end up spending more money on compute. OpenAI just signed a huge $11.9 billion five year contract with one of their cloud providers.
In other words, don’t expect ChatGPT to get cheaper, but you won’t have to wait as long for GPT-5.
New in the Universe: What’s a context window?
Let’s close with a favorite term we added to the Technically universe of technical concepts this month — the context window:
A context window is how much data an AI model can hold in memory at once; it's like the RAM of AI models.
In the context of Large Language Models, the context window dictates how long your prompt can be, what kinds of documents you can include, and how much the model "remembers" from earlier in the conversation.
Write for Technically
We’re seeking an analyst to help expand Technically’s coverage of software companies within categories like AI + Analytics, Cloud Infrastructure, Data Stores + DevOps.
If your father always wanted you to be a banker, but you have a writer inside you waiting to get out, we’d love to meet you — just hit reply.
Coming up next month
Thanks for joining us on this inaugural Technically Monthly. Stay tuned for the next one, which will feature upcoming posts on:
How do you train an AI model?
What’s CSS?
How can AI use websites?
Love the new format!