> log #04 — what AI headshots tell us about the new economy
An image is worth a trillion chips: decoding the economic signals behind better AI headshots.
In 2023, I tried an AI headshot generator for the first time. The results were… okay. A few shots kind of looked like me, but most looked like a slightly melted version of my face, posing awkwardly in an uncanny valley.
I tried again in 2025 using the same tool, HeadshotPro — and the difference was jaw-dropping. This time, I got polished, studio-grade portraits that I actually used on LinkedIn and in professional branding. But this shift isn’t just about better photos.
It’s a window into the entire AI-driven economy — from the hardware that powers it to the stocks that benefit, to the data centers rising in towns you’ve never heard of.
Let’s break it down.
1. Better AI Models → Better Results → More Demand
In 2023, most image tools used Stable Diffusion 1.5, an open-source model that could turn text into images — but struggled with faces, lighting, and realism. That’s because early diffusion models weren’t trained on high-resolution human portrait datasets, and lacked the spatial awareness to render convincing facial geometry or lighting.
By 2025, most professional-grade tools (like HeadshotPro, Secta, and ProPhotos AI) have moved to:
Stable Diffusion XL (SDXL): Trained on higher-resolution data, with much more detailed facial modeling
Diffusion Transformers (DiT): A newer architecture that brings transformer logic (used in ChatGPT) into image generation
// econ impact:
These newer models are compute-hungry. You need high-end GPUs and cloud-scale infrastructure to run them efficiently — and that’s where NVIDIA comes in.
NVIDIA’s flagship AI chip, the H100, is now the most in-demand piece of silicon in the world. AI companies lease clusters of these chips to run image generation, video synthesis, LLMs, and more.
In 2023, NVIDIA’s market cap passed $1 trillion. In 2025, it’s hovering near $3 trillion. That rise is fueled not just by hype — but by real, sustained demand for high-compute inference across image and language models.
If you’re wondering why headshots matter: they’re a perfect example of consumer AI demand that requires industrial-scale backend power.
Personalization Is Easier Now → More Use Cases → More Labor Augmentation
The 2023 approach to personalization used DreamBooth, a fine-tuning method that retrained the entire AI model on 10–20 selfies. This was expensive, inconsistent, and slow. Most of the shots looked like you had been run through a cosplay filter.
In 2025, most tools use LoRA (Low-Rank Adaptation) or FaceID-based embedding alignment:
These allow for fast personalization using a smaller training footprint
You get better face consistency across dozens of poses and outfits
The results are usable, not just fun
This shift opened the door for wider business adoption: companies now use AI headshots for employee directories, professional branding, and even customer avatars.
// econ impact:
This is a form of labor augmentation — creative freelancers, marketing teams, recruiters, and even small business owners are replacing one-time professional photography costs with $20–$50 AI subscriptions.
It doesn’t eliminate jobs — it reshapes them. There’s rising demand for people who know how to guide AI, prompt well, and post-process creatively.
The market for “AI-native” creative professionals is growing. New job titles include prompt engineers, synthetic stylists, and AI brand consultants.
3. Faster Compute → Lower Margins, Higher Throughput → Infrastructure Gold Rush
Behind every AI headshot you see is a network of cloud GPUs running inference jobs across huge datasets and models.
In 2023, most inference was done on NVIDIA A100s — powerful, but expensive. In 2025, most companies have upgraded to H100s, and many are already planning for Blackwell B100s (NVIDIA’s next-gen chips).
H100s are not just faster — they’re optimized for transformer models and generative workloads. One H100 can process nearly 5x the throughput of an A100 in some tasks, and with lower energy per output.
// econ impact:
This has triggered a buildout of AI infrastructure at scale. New data centers — often called “AI factories” — are being constructed across the U.S., especially in areas with cheap energy (like Iowa, Utah, and parts of Texas).
Companies like CoreWeave, Lambda Labs, Microsoft Azure, and Google Cloud are all racing to provide capacity.
This infrastructure boom is creating a new kind of industrial economy — not based on steel or oil, but on compute, cooling, and connectivity.
Even if all you want is a new LinkedIn photo, that request is tied to terawatts of energy, silicon fabrication plants in Taiwan, and multi-billion-dollar infrastructure leases.
Better Post-Processing → Higher Trust → Stronger Consumer Adoption
In 2023, AI headshots still looked… like AI. They often had strange shadows, dead eyes, and awkward hair edges.
By 2025, tools now filter and clean every image using:
Aesthetic scoring models (to rank what looks good to humans)
Face integrity filters (to catch distortions)
Enhancement tools like CodeFormer or GFPGAN to sharpen and clean final images
The result? A product people actually trust.
// econ impact:
This trust leads to higher conversion: people buy these photos, use them professionally, and share them — which fuels organic growth loops.
That in turn drives more small-scale AI startups that focus on vertical tools — headshots, e-commerce product photos, dating profile shots, real estate staging, and more.
We’re witnessing a fragmentation of the AI SaaS market: instead of one general-purpose model, we’re getting thousands of niche tools built on top of shared backends (usually OpenAI or Stability + NVIDIA).
Want a new pfp? 📸
> closing loop:
Short-Term Economic Forecast for End of 2025
Global Data Center Growth
Expect 10 GW of new hyperscale and colo-powered capacity to break ground globally in 2025, with 7 GW coming online by year’s end (≈ $170 billion in asset value).
U.S. Data Center Spending
Infrastructure investment for AI-capable data centers in the U.S. will account for a significant share of the $1.8 trillion projected global build-out by 2030.
Power Demand Surge
U.S. data centers are expected to consume up to 9% of national electricity by 2030. Globally, data center energy usage is projected to nearly double to 945 TWh by then.
Job Creation: Short-Term vs Long-Term
– Example: The CoreWeave 100 MW PA site (~300 MW planned) is estimated to generate 600 construction jobs plus ~175 permanent roles.
– But full-time operational staff tend to be ~50–150 per site, leading to concerns over low long-term job density.
Tech Workforce Shift
– Engineers and facility technicians are being pushed to learn AI-specialized skills just to manage the infrastructure.
Macro Capex & GDP Impact
– AI-related capital expenditures (capex) may hit ~2% of U.S. GDP by end of 2025 — that’s around $550 billion with knock-on effects in energy, real estate, and labor.
What This Means by End of 2025
Construction spike — tens of thousands of jobs building large-scale AI data centers.
Operational restraint — hundreds, not thousands, of long-term roles per site.
Power grid stress — massive new energy draw requiring upgrades & sustainability planning.
Local infrastructure boom — real estate demand, tax revenues, and utility projects pouring into rural & secondary markets.
Skill-based hiring surge — demand for data center engineers and AI-savvy operational staff grows faster than university pipelines .
🧾 TL;DR
By the end of 2025, AI and data center expansion will:
Break ground on 7–10 GW of new capacity globally,
Create 600+ construction gigs per site, but only ~100–150 permanent roles,
Drive capex spending equal to ~2% of U.S. GDP,
Force utilities to scramble to support growing power loads,
And tip the labor market toward compute infrastructure skills.