BotW is closer to the 1986 Legend of Zelda than any other Zelda game


In the mid-1980s, in Kyoto, Nintendo’s Shigeru Miyamoto, Takashi Tezuka, and their team were working on two game designs in parallel that were intended to be two sides of the same coin. One would be linear, an obstacle course, a headlong rush toward a goal. The other would be nonlinear, a mysterious maze, an unpredictable world of exploration. As opposed to the arcade games that defined the medium at that point, both would deemphasize the player’s skill in favor of an immersive experience with a beginning and an end. Miyamoto would carefully consider each new gameplay idea from the team and assign it to one game or the other, probably unaware that he was defining not just two legendary series, but some of the foundational precepts of video game design. The first game was Super Mario Bros. The second game was The Legend of Zelda.

Super Mario Bros. was a phenomenon and remains one of the best-selling games of all time. The Legend of Zelda was a hit, but sold a fraction as many copies. Perhaps it was inevitable that Zelda would later become a little more like its more popular twin. Sequels kept the spirit of adventure alive, but used meticulous gear-gating systems — essentially a series of locks and keys — to enforce order on the player’s progress through their labyrinthine worlds. They were probably better games for it, and certainly easier to enjoy.

2017’s Breath of the Wild was a shock to that system: an incredibly free-form reinvention of role-playing and open-world conventions, and a return to Zelda’s original nonlinear philosophy. Ironically, it would outsell all of its predecessors by a huge margin, and also become the first Zelda to outsell a contemporary Mario game. The world was finally ready for the kind of adventure Miyamoto and Tezuka had imagined 31 years prior.

Even knowing all this context as I do — and having written about Zelda games for 20 years, and played them for even longer — it has been an equivalent shock to properly play The Legend of Zelda for the first time. It’s a very old game now: difficult, inscrutable, and minimalist by modern standards. Honestly, it’s nowhere near as much fun to play as 1991’s A Link to the Past, never mind later entries. But it’s also game design genius in the raw. The Legend of Zelda is a stunningly brave and unencumbered vision of what a playable adventure could be. It’s both deeply familiar as the father of all Zelda games, and almost alien in its distance from what they went on to become — until Breath of the Wild, that is.

The first thing you notice is that, while most Zelda games (Breath of the Wild included) consciously locate themselves in a tradition — a place with a history, usually one that repeats itself — The Legend of Zelda strips away all context. There’s a perfunctory setup: Princess Zelda has hidden the eight parts of the Triforce of Wisdom in a series of dungeons, and she needs a hero to gather them and defeat her captor, Ganon. Enter Link, who in this game is just a guy, standing in a clearing, figuring out what to do next. He doesn’t even have a sword.

On this first adventure, nothing has been foretold or forgotten yet, and everything is new. There’s a cave over there: What’s in it? An old man, with a wooden sword and a warning: “It’s dangerous to go alone.” But there’s no choice, so alone Link must go. Monsters are everywhere, skittering about in treacherous patterns as personalized as they are unpredictable. Which way to the next piece of Triforce? Into the thorn bushes, among the bare rocks, along the beach with its murmuring tides? This is a world without waypoints, signposts, or even names. It’s a mysterious iconography, a living map of which you can only ever see a tiny portion.

In The Legend of Zelda, the first dungeon you discover won’t necessarily be the game’s first “level.” You might have jumped ahead to the third or fourth. Equally, if you set off and explore, and take great care, you might find your way to powerful items, intended for use much later in the game, within your first hour. There are riches scattered around, if you know where to look. Buy a candle and burn the right bush, and you can climb underground to meet a secretive Moblin who’ll bestow you with an amount of Rupees that would take hours to grind out. Now you can buy potions, arrows for a bow you don’t have yet, a blue ring that cuts the damage you take by half.

Even by most modern standards, this is daringly open design. You can’t wander straight into the end boss’s lair, but you can tumble into great danger or power up Link to the extent the game feels almost broken. The few gates that Nintendo did place in the game’s structure feel like genuine mysteries, arising organically from the landscape — a feeling that Zelda’s designers would become expert at re-creating, but in a way that, over time, became expected and almost ceremonial. Few Zelda fans, I imagine, wanted to let the ceremonies of that more linear evolution go, but diminishing returns were in effect, and Breath of the Wild’s developers were right to choose to move past them. What they sought to recapture (and did) was something that even the very best Zelda games had missed in the intervening decades: The Legend of Zelda’s Hyrule feels like an actual, unexplored wilderness, or like the countryside that Miyamoto used to explore in his youth, without a map, discovering landmarks as if he were the first to ever go there.

Even by most modern standards, this is daringly open design

In contrast to its untamed overworld, the first game’s dungeons, which would later become the Zelda series’ most intricate and exacting puzzles, feel a little more contained and more surmountable, even when they’re at their most difficult. That probably wasn’t true at the time of its release. These are the first words in a language that has since been greatly elaborated on, and solving their riddles will come as second nature to any seasoned Zelda player. The first game’s basic, ferocious 8-bit combat presents a challenge, but as technically rudimentary as the game is, it’s never less than razor-sharp, responsive, and fair — as slickly playable as you would expect from the team that had shipped Super Mario Bros. just months previously. (I had to spam the snapshot and rewind features of the Nintendo Switch Online version to beat it, though.)

The Legend of Zelda is so sparing with its clues, so gnomic in its design, so jealous of its secrets, that a new player attempting to solve it without help will probably get stumped. If you do look for assistance, don’t feel bad: Miyamoto may not have foreseen online walkthroughs, but he always intended players of the game to talk to one another, share secrets, and collaborate, getting to the end through communal effort.

Unless you’re playing The Legend of Zelda in some kind of cultural vacuum, it’s no longer possible to experience the game in all the formidable mystery it possessed in 1986. It’s a known quantity now, and has passed into a modern version of folk memory: a tale already told, a map already drawn. In some ways, that’s a fitting end for this early masterpiece. And we can still look down on this teeming, savage little microcosm and wonder at it: a game so free and so far ahead of its time that it took its own makers 30 years to catch up.

Free Church Live Streaming & AV Event


Join us February 24–26 at 3PM EST for WorshipStreamTV, a free 3-day online event dedicated to modern church live streaming, video production, and AV workflows for houses of worship.

If you’re looking to upgrade aging PTZ cameras, improve your IMAG setup, expand your online ministry, or simplify volunteer-friendly workflows, this event is designed specifically for you.

👉 Free Registration: https://worshipstream.tv/

Why WorshipStreamTV Matters for Churches

Church live streaming and in-house video production have become essential tools for ministry. Whether you’re serving a small congregation or a large multi-campus church, having reliable, scalable, and easy-to-use technology is critical.

WorshipStreamTV focuses on helping churches:

  • Upgrade outdated PTZ systems
  • Improve IMAG (Image Magnification) quality
  • Implement AV-over-IP workflows
  • Grow online viewership
  • Simplify volunteer-operated production systems
  • Integrate LED video walls
  • Produce high-quality pre-recorded and live content

This event is built for church tech directors, media volunteers, pastors, and worship leaders who want practical, real-world solutions.

Hear from Leading Church Production Brands

Learn directly from industry leaders shaping modern worship production:

  • BirdDog
  • Kiloview
  • Panasonic
  • PTZOptics
  • QuickLink
  • Vizrt

These sessions will cover real-world workflows and tools used in churches across the country — from small sanctuaries to large auditoriums.

Topics We’ll Cover

WorshipStreamTV is packed with actionable sessions on:

  • Church IMAG & Multi-Camera Production: Learn how to improve in-room engagement with high-quality IMAG systems using PTZ cameras and broadcast-style switching.
  • AV-over-IP & NDI Workflows: Discover how IP-based production simplifies cable runs, increases flexibility, and allows scalable expansion.
  • Growing Your Church Livestream: Understand strategies for increasing engagement on YouTube, Facebook, and church platforms while maintaining professional production standards.
  • Volunteer-Friendly Control Systems: Explore intuitive switching systems and control surfaces that empower volunteers to operate confidently.
  • LED Video Walls for Churches: See how LED walls enhance worship environments and create immersive experiences.
  • Produced Content & Hybrid Ministry: Learn how to integrate pre-produced content with live worship for a seamless in-person and online experience.

Prizes — Register to Win!

When you register for WorshipStreamTV, you’ll automatically be entered to win:

  • PTZOptics Studio 4K 12X Zoom Camera
  • Elgato Stream Deck MK.2
  • A $100 Videoguys Gift Card
  • No purchase required — just register and attend!

Event Details

📅 February 24–26
⏰ 3PM EST each day
🎟 Free Online Event

👉 Reserve Your Spot Now: https://worshipstream.tv/

One last call: Samsung’s One UI 8.5 Beta 5 shows up before the Galaxy S26 takes the stage


What you need to know

  • Samsung has reportedly started rolling out its One UI 8.5 Beta 5 to Galaxy S25 users in the U.S.
  • This patch is quite light, as the patch notes mainly detail an “applied Bixby version update,” which could be the major rework we’ve seen rumored before.
  • Samsung’s Galaxy Unpacked takes place next week on February 25, 2026 at 1 pm ET/10 am PT.

We are less than a week away from Samsung’s flurry of next-gen announcements, and it’s looking to make sure its next software is ready for it.

Samsung’s One UI 8.5 Beta 5 was spotted rolling out to eligible Galaxy S25 users by SammyGuru. Reports state that the patch is slowly becoming available for users who’ve enrolled in the beta program in the U.S. The changelog for the patch is quite light, as Samsung labels this as One UI 8.5 vZZAN. More than that, the update applies “Bixby version update.”

Hello Kawaii Food Coloring Pages (Cute Snacks, Drinks, Fruit, Veggies, & Desserts)


This is gonna sound crazy, I know, but I LOVE food! Here’s the thing…when we get to Heaven, I highly doubt there’s gonna be a Big Mac for us to eat, you know what I mean? 😆

And since Heaven is for eternity, I need to eat and enjoy food NOW, while I’m here on Earth.

To savor it.

To enjoy it.

And boy, do I enjoy it! 😂

I love to cook. I love to try different foods. I love to eat my favorite foods. I’m not one of those people who eat a tiny bit just enough to survive and are good to go. I don’t gorge myself either, but I like to enjoy the food.

That view of food, probably isn’t the most healthiest of choices for me here on Earth, but I gotta say, I’m enjoying it while I can. Before I get old and all I can eat is a blender full of steak and mashed potatoes 🤢 (that’s a legit thing, my ex-husband works in senior assisted living places for his job) or rice cakes. 😆

No. I’m good. I want to enjoy food right now while I can!

I like food.

Maybe you like food too, but you don’t think of it to that extent, I get it, I’m weird…always have been, 🤪 I accept it. 😆

But maybe you’d still love to color some really cute, AMAZINGLY CUTE, Hello Kawaii Food Coloring Pages!

Maybe you’re trying to get your kids to like vegetables more. #mombattles

Give them the fruits and veggies coloring sheets to encourage them. ❤️

Maybe you’re a bakery who makes cupcakes that blows people’s socks off. Put the Hello Cupcake coloring page in your bakery with a cup of crayons for kids to color and promote your yummy baked treats. 🧁

Maybe you made a cupcake for a friend and wanna give them something extra in your gift. Grab the Hello Cupcake coloring sheet and put it in the envelope with a cupcake sitting on top.

Maybe you just wanna blow off some steam and relax, treat yo’ self and color. 🤗 No probs!

Whatever you do with these Hello Food Coloring Sheets, I’m sure it’ll be amazing! They are just SUPER adorable!!!! 🥰

Cute Food Coloring Pages To Print And Color

These kawaii food coloring pages are so adorable, I wanted to make all different options for you so you can pick and choose which one(s) you want to color, or color them all! 🎉

Cute Fast Food Coloring Sheet

This page is the full on hello food mix with cute fast food, sweets, and all kinds of adorable little bites.

In this hello food coloring page, you’ll see things like a popsicle, pretzel, french fries, ramen cup, popcorn, burger with a bow, sandwich, donut, chocolate bar, cherries, avocado, coffee cup, pizza slice, pineapple, pie, bubble tea, taco, burger, strawberry, oranges, sushi, waffle with whipped cream, noodle takeout box, corn dog, lemon slice, banana, croissant, dumpling, watermelon slice, baguette, toast, peas in a pod.

Cute kawaii food coloring page featuring smiling snacks and desserts like burgers, fries, donuts, waffles, ice cream, and fruit in a fun black-and-white doodle style.Cute kawaii food coloring page featuring smiling snacks and desserts like burgers, fries, donuts, waffles, ice cream, and fruit in a fun black-and-white doodle style.

Download the Hello Food Coloring Page Kawaii Style Printable here. 

 

Adorable Cupcake Coloring Sheet

Cupcake heaven. If you love frosting swirls and sweet little faces, this page is going to be your favorite.

You’ll see things like a cupcake with almond, cupcake with a cherry, mini swirl cupcake, cupcake with bow, unicorn cupcake, cupcake with blueberries, cupcake with strawberry, bunny cupcake, swirl cupcake, fluffy cupcake, cupcake with sprinkles and wafer stick, muffin, cupcake with strawberry and wafer sticks, rainbow cupcake, cupcake with strawberry slice, winking cupcake, chick cupcake, cupcake with heart topper.

Hello Cupcake kawaii coloring page filled with cute smiling cupcakes, sprinkles, hearts, and playful dessert doodles in black-and-white line art.Hello Cupcake kawaii coloring page filled with cute smiling cupcakes, sprinkles, hearts, and playful dessert doodles in black-and-white line art.

Download the Hello Cupcake Kawaii Food Colorable Sheet PDF here. 

 

Yummy Dessert Coloring Sheet

Desserts only. Cakes, ice cream, donuts, and all the sweet things with the cutest faces.

In this Kawaii dessert coloring page, you’ll see things like a ice cream milkshake with whip cream, soft serve ice cream cone, pie slice with whipped cream, pancakes, layer cake slice, pie slice, ice cream bowl, ice cream sundae, cake with strawberry, animal scoop ice cream cone, cotton candy, birthday cake, parfait, unicorn donut, small sundae, bunny ice cream cone, cotton candy character, jelly, pudding, roll cake, popsicle, whole pie, double scoop ice cream cone, strawberry cake, ice cream cone with cherry, cake slice with fork, heart ice cream cone, bitten popsicle, parfait glass, soft serve cone.

Hello Desserts kawaii coloring page featuring cute smiling cakes, ice cream cones, donuts, and sweet treats in playful black-and-white line art.Hello Desserts kawaii coloring page featuring cute smiling cakes, ice cream cones, donuts, and sweet treats in playful black-and-white line art.

Download the Printable Hello Desserts Kawaii Food Coloring Sheet here. 

 

Happy Fruit Coloring Printable

Bright, happy, and super cute. This fruit page is perfect when you want something cheerful and light.

In this fruit coloring page printable, you’ll see things like a half of a watermelon, orange, watermelon wedge, peach, lemon, strawberry, avocado, blueberries, bitten watermelon slice, apple, kiwi slice, pineapple, peeled banana, grapes, watermelon popsicle, banana, jam jar, dragon fruit slice, cherries, tomato, mangosteen, candy apple, watermelon slice, persimmon, trio of berries.

Hello Fruit kawaii coloring page filled with cute smiling fruits like watermelon, strawberries, apples, bananas, and citrus in fun black-and-white line art.Hello Fruit kawaii coloring page filled with cute smiling fruits like watermelon, strawberries, apples, bananas, and citrus in fun black-and-white line art.

Download the PDF Hello Fruit Kawaii Food Coloring Page here. 

 

Kawaii Drinks and Boba Tea Coloring Sheet

All the cozy drink vibes, from boba to coffee to cute little mugs. This one is so relaxing to color.

In this kawaii drinks coloring page, you’ll see things like a juice glass with orange slice, glass with straw, mug with heart, small cup, punch glass, coconut drink, hot cocoa with marshmallows, iced drink jar, soda glass with lime, mug on coaster, coffee with whipped cream, bubble tea, takeout coffee cup, teapot, lemonade glass, latte cup, soda bottle, berry drink, steaming bowl, tilted bubble tea, mason jar drink, tall glass with straw, bear mug, milk bottle, latte art, frappe, floral tea cup, thermos, juice box, iced coffee, polka dot mug, espresso cup, cocktail glass with strawberry, fruit pitcher, jar with ice.

Hello Drinks kawaii coloring page featuring cute smiling mugs, coffee, tea, smoothies, bubble tea, and cozy drink doodles in black-and-white line art.Hello Drinks kawaii coloring page featuring cute smiling mugs, coffee, tea, smoothies, bubble tea, and cozy drink doodles in black-and-white line art.

Download the Hello Drinks Food Coloring Page Printable here. 

 

Kawaii Snacks and Candy Coloring Page

This kawaii snacks coloring page is loaded with cute candies, cookies, and all the little snacks that make you smile while you color.

In this cute snack coloring sheet, you’ll see things like cotton candy, wrapped candy, striped popsicle, bunny ice cream cone, bitten popsicle, unicorn donut, bunny popsicle, swirl lollipop, candy corn, boba tea, snack bag, round lollipop, peach soda, pinwheel lollipop, pretzel, ring pop, hard candy, chocolate bar, chocolate chip cookie, ice cream sundae, gumball machine, ice cream cone, bag of candies, heart lollipop, donut.
Hello Snacks kawaii coloring page with cute smiling candies, donuts, popsicles, cookies, pretzels, and fun snack doodles in black-and-white line art.Hello Snacks kawaii coloring page with cute smiling candies, donuts, popsicles, cookies, pretzels, and fun snack doodles in black-and-white line art.

Download the Hello Snacks Kawaii Food Coloring Sheet PDF here. 

 

Cute Vegetable Coloring Page

Veggies, but made adorable. This page is surprisingly fun, and it has that playful garden vibe.

In this veggie coloring pages printable, you’ll see things like cabbage, a bowl of veggies, chili pepper, cauliflower, carrot, artichoke, broccoli, eggplant, garlic, cucumber slices, tomatillo, lettuce, peas in a pod, onion, corn on the cob, radish, and kabocha squash.

Hello Veggies kawaii coloring page with cute smiling vegetables like carrots, broccoli, corn, peas, tomatoes, and leafy greens in black-and-white line art.Hello Veggies kawaii coloring page with cute smiling vegetables like carrots, broccoli, corn, peas, tomatoes, and leafy greens in black-and-white line art.

Download the Printable Kawaii Hello Veggies Food Coloring Sheet here. 

 

What sets your printables apart?

• Other printables are either functional OR cute. I bridge the gap between the two. Since I’m a single mom and don’t have time to mess around and am a kid at heart, I love something that actually propels me forward and gets the job done quickly, while still having that super cute vibe I’m known for.
• I’m a graphic designer FIRST. This means my printables, whether free or paid for, will always be of the highest quality and print crisp and clear on any printer.
• I love what I do and I do the work myself. I’m not hiring these out. This is my life, my hobby, my passion, my heart.
• My fans are obsessed with my work. To date, I’ve made over $23M on printables alone (organically) in the past 8 years.

Is this printable free to download?

Yes! Except for some large binders, all printables on the blog are 100% free for personal use. I make these because I genuinely want you to have something cute and helpful right now without having to spend a dime. You deserve a life that feels more peaceful, more organized, and more you, and if a simple printable can make your day feel lighter, I am so happy to give that to you.

Can I use these printables for my classroom or church group?

Absolutely! You’re welcome to print as many copies as you need for your own students or church groups. I only ask that you do not edit them in any way (leaving the full copyright line in tact) or sell them or host the digital files on other websites. Please always link back to this post to share with others.

What is the best way to print these high-quality printables?

For best results, I recommend printing on 8.5″x11″ white paper. These are professionally designed as high-resolution PDFs so they will stay crisp and clear, making them perfect for any printer.

Password managers’ promise that they can’t see your vaults isn’t always true



Over the past 15 years, password managers have grown from a niche security tool used by the technology savvy into an indispensable security tool for the masses, with an estimated 94 million US adults—or roughly 36 percent of them—having adopted them. They store not only passwords for pension, financial, and email accounts, but also cryptocurrency credentials, payment card numbers, and other sensitive data.

All eight of the top password managers have adopted the term “zero knowledge” to describe the complex encryption system they use to protect the data vaults that users store on their servers. The definitions vary slightly from vendor to vendor, but they generally boil down to one bold assurance: that there is no way for malicious insiders or hackers who manage to compromise the cloud infrastructure to steal vaults or data stored in them. These promises make sense, given previous breaches of LastPass and the reasonable expectation that state-level hackers have both the motive and capability to obtain password vaults belonging to high-value targets.

A bold assurance debunked

Typical of these claims are those made by Bitwarden, Dashlane, and LastPass, which together are used by roughly 60 million people. Bitwarden, for example, says that “not even the team at Bitwarden can read your data (even if we wanted to).” Dashlane, meanwhile, says that without a user’s master password, “malicious actors can’t steal the information, even if Dashlane’s servers are compromised.” LastPass says that no one can access the “data stored in your LastPass vault, except you (not even LastPass).”

New research shows that these claims aren’t true in all cases, particularly when account recovery is in place or password managers are set to share vaults or organize users into groups. The researchers reverse-engineered or closely analyzed Bitwarden, Dashlane, and LastPass and identified ways that someone with control over the server—either administrative or the result of a compromise—can, in fact, steal data and, in some cases, entire vaults. The researchers also devised other attacks that can weaken the encryption to the point that ciphertext can be converted to plaintext.

Following layoffs, Highguard devs Wildlight announce it’s getting a new airship base and door-opening dart gun


When the website of Highguard suddenly began blaring that it was “unavailable” not long after news broke of layoffs at developer Wildlight, you could have been forgiven for thinking a Concord-style pulling offline might be in the shooter’s near future. That doesn’t look to be the case, though, with the studio having just announced a couple of new additions set to arrive in the game this week.

Continue reading “Following layoffs, Highguard devs Wildlight announce it’s getting a new airship base and door-opening dart gun”

Custom Agents in Visual Studio: Built in and Build-Your-Own agents


Agents in Visual Studio now go beyond a single general-purpose assistant. We’re shipping a set of curated preset agents that tap into deep IDE capabilities; debugging, profiling, testing alongside a framework for building your own custom agents tailored to how your team works.

Each preset agent is designed around a specific developer workflow and integrates with Visual Studio’s native tooling in ways that a generic assistant can’t.

  • Debugger – Goes beyond “read the error message.” Uses your call stacks, variable state, and diagnostic tools to walk through error diagnosis systematically across your solution.
  • Profiler – Connects to Visual Studio’s profiling infrastructure to identify bottlenecks and suggest targeted optimizations grounded in your codebase, not generic advice.
  • Test – (when solution is loaded) Generates unit tests tuned to your project’s framework and patterns, not boilerplate that your CI will reject.
  • Modernize (.NET and C++ only) -Framework and dependency upgrades with awareness of your actual project graph. Flags breaking changes, generates migration code, and follows your existing patterns.

Access them through the agent picker in the chat panel or using ‘@’ in chat.

The presets cover workflows we think matter most, but your team knows your workflow better than we do. Custom agents let you build your own using the same foundation—workspace awareness, code understanding, tools accessed by your prompts, your preferred model, and your tools.

Where it gets powerful is MCP. You can connect custom agents to external knowledge sources internal documentation, design systems, APIs, and databases so the agent isn’t limited to what’s in your repo.

A few patterns we’re seeing from teams:

  • Code review that checks PRs against your actual conventions, connected via MCP to your style guide or ADR repository
  • Design system enforcement connected to your Figma files or component libraries to catch UI drift before it ships
  • Planning helps you think through a feature or task before any code is written. Gathers requirements, asks clarifying questions, and builds out a plan that you can hand off

The awesome-copilot repo has community-contributed agent configurations you can use as starting points.

Get started

Custom agents are defined as .agent.md files in your repository’s .github/agents/ folder:

your-repo/
└── .github/
    └── agents/
        └── code-reviewer.agent.md

A few things to note:

  • This is a preview feature; the format of these files may change over to support different capabilities
  • If you don’t specify a model, the agent uses whatever is selected in the model picker
  • Tool names vary across GitHub Copilot platforms- check the tools available in Visual Studio specifically to make sure your agent works as expected
  • Configurations from the awesome-copilot repo are a great starting point, but verify tool names before using them in VS

Tell us what you’re building

Share your configurations in the awesome-copilot repo or file feedback here.

How to Protect Your Home Business During Divorce






How to Protect Your Home Business During Divorce























































Multi-GPU vs Single-GPU Scaling economics


Introduction—Why scale economics matter more than ever

The modern AI boom is powered by one thing: compute. Whether you’re fine‑tuning a vision model for edge deployment or running a large language model (LLM) in the cloud, your ability to deliver value hinges on access to GPU cycles and the economics of scaling. In 2026 the landscape feels like an arms race. Analysts expect the market for high‑bandwidth memory (HBM) to triple between 2025 and 2028. Lead times for data‑center GPUs stretch over six months. Meanwhile, costs lurk everywhere—from underutilised cards to network egress fees and compliance overhead.

This article isn’t another shallow listicle. Instead, it cuts through the hype to explain why GPU costs explode as AI products scale, how to decide between single‑ and multi‑GPU setups, and when alternative hardware makes sense. We’ll introduce original frameworks—GPU Economics Stack and Scale‑Right Decision Tree—to help your team make confident, financially sound decisions. Throughout, we integrate Clarifai’s compute orchestration and model‑inference capabilities naturally, showing how a modern AI platform can tame costs without sacrificing performance.

Quick digest

  • What drives costs? Scarcity in HBM and advanced packaging; super‑linear scaling of compute; hidden operational overhead.
  • When do single GPUs suffice? Prototyping, small models and latency‑sensitive workloads with limited context.
  • Why choose multi‑GPU? Large models exceeding single‑GPU memory; faster throughput; better utilisation when orchestrated well.
  • How to optimise? Rightsize models, apply quantisation, adopt FinOps practices, and leverage orchestration platforms like Clarifai’s to pool resources.
  • What’s ahead? DePIN networks, photonic chips and AI‑native FinOps promise new cost curves. Staying agile is key.

GPU Supply & Pricing Dynamics—Why are GPUs expensive?

Context: scarcity, not speculation

A core economic reality of 2026 is that demand outstrips supply. Data‑centre GPUs rely on high‑bandwidth memory stacks and advanced packaging technologies like CoWoS. Consumer DDR5 kits that cost US$90 in 2025 now retail at over US$240, and lead times have stretched beyond twenty weeks. Data‑centre accelerators monopolise roughly 70 % of global memory supply, leaving gamers and researchers waiting in line. It’s not that manufacturers are asleep at the wheel; building new HBM factories or 2.5‑D packaging lines takes years. Suppliers prioritise hyperscalers because a single rack of H100 cards priced at US$25 K–US$40 K each can generate over US$400 K in revenue.

The result is predictable: prices soar. Renting a high‑end GPU on cloud providers costs between US$2 and US$10 per hour. Buying a single H100 card costs US$25 K–US$40 K, and an eight‑GPU server can exceed US$400 K. Even mid‑tier cards like an RTX 4090 cost around US$1,200 to buy and US$0.18 per hour to rent on marketplace platforms. Supply scarcity also creates time costs: companies cannot immediately secure cards even when they can pay, because chip vendors require multi‑year contracts. Late deliveries delay model training and product launches, turning time into an opportunity cost.

Operational reality: capex, opex and break‑even math

AI teams face a fundamental decision: own or rent. Owning hardware (capex) means large upfront capital but gives full control and avoids price spikes. Renting (opex) offers flexibility and scales with usage but can be expensive if you run GPUs continuously. A practical break‑even analysis shows that for a single RTX 4090 build (~US$2,200 plus ~US$770 per year in electricity), renting at US$0.18/hr is cheaper unless you run it more than 4–6 hours daily over two years. For high‑end clusters, a true cost of US$8–US$15/hr per GPU emerges once you include power distribution upgrades (US$10 K–US$50 K), cooling (US$15 K–US$100 K) and operational overhead.

To help navigate this, consider the Capex vs Opex Decision Matrix:

  • Utilisation < 4 h/day: Rent. Cloud or marketplace GPUs minimise idle costs and let you choose hardware per job.
  • Utilisation 4–6 h/day for > 18 months: Buy single cards. You’ll break even in the second year, provided you maintain usage.
  • Multi‑GPU or high‑VRAM jobs: Rent. The capital outlay for on‑prem multi‑GPU rigs is steep and hardware depreciates quickly.
  • Baseline capacity + bursts: Hybrid. Own a small workstation for experiments, rent cloud GPUs for big jobs. This is how many Clarifai customers operate today.

elasticity and rationing

Scarcity isn’t just about price—it’s about elasticity. Even if your budget allows expensive GPUs, the supply chain won’t magically produce more chips on your schedule. The triple‑constraint (HBM shortages, advanced packaging and supplier prioritisation) means the market remains tight until at least late 2026. Because supply cannot meet exponential demand, vendors ration units to hyperscalers, leaving smaller teams to scour spot markets. The rational response is to optimise demand: right‑size models, adopt efficient algorithms, and look beyond GPUs.

What this does NOT solve

Hoping that prices will revert to pre‑2022 levels is wishful thinking. Even as new GPUs like Nvidia H200 or AMD MI400 ship later in 2026, supply constraints and memory shortages persist. And buying hardware doesn’t absolve you of hidden costs; power, cooling and networking can easily double or triple your spend.

Expert insights

  • Clarifai perspective: Hyperscalers lock in supply through multi‑year contracts while smaller teams are forced to rent, creating a two‑tier market.
  • Market projections: The data‑centre GPU market is forecast to grow from US$16.94 B in 2024 to US$192.68 B by 2034.
  • Hidden costs: Jarvislabs analysts warn that purchasing an H100 card is only the beginning; facility upgrades and operations can double costs.

Quick summary

Question – Why are GPUs so expensive today?

Summary – Scarcity in high‑bandwidth memory and advanced packaging, combined with prioritisation for hyperscale buyers, drives up prices and stretches lead times. Owning hardware makes sense only at high utilisation; renting is generally cheaper under 6 hours/day. Hidden costs such as power, cooling and networking must be included.

Mathematical & Memory Scaling – When single GPUs hit a wall

Context: super‑linear scaling and memory limits

Transformer‑based models don’t scale linearly. Inference cost is roughly 2 × n × p FLOPs, and training cost is ~6 × p FLOPs per token. Doubling parameters or context window multiplies FLOPs more than fourfold. Memory consumption follows: a practical guideline is ~16 GB VRAM per billion parameters. That means fine‑tuning a 70‑billion‑parameter model demands over 1.1 TB of GPU memory, clearly beyond a single H100 card. As context windows expand from 32 K to 128 K tokens, the key/value cache triple in size, further squeezing VRAM.

Operational strategies: parallelism choices

Once you hit that memory wall, you must distribute your workload. There are three primary strategies:

  1. Data parallelism: Replicate the model on multiple GPUs and split the batch. This scales nearly linearly but duplicates model memory, so it’s suitable when your model fits in a single GPU’s memory but your dataset is large.
  2. Model parallelism: Partition the model’s layers across GPUs. This allows training models that otherwise wouldn’t fit, at the cost of extra communication to synchronise activations and gradients.
  3. Pipeline parallelism: Stages of the model are executed sequentially across GPUs. This keeps all devices busy by overlapping forward and backward passes.

Hybrid approaches combine these methods to balance memory, communication and throughput. Frameworks like PyTorch Distributed, Megatron‑LM or Clarifai’s training orchestration tools support these paradigms.

when splitting becomes mandatory

If your model’s parameter count × 16 GB > available VRAM, model parallelism or pipeline parallelism is non‑negotiable. For example, a 13 B model needs ~208 GB of VRAM; even an H100 with 80 GB cannot host it, so splitting across two or three cards is required. The PDLP algorithm demonstrates that careful grid partitioning yields substantial speedups with minimal communication overhead. However, just adding more GPUs doesn’t guarantee linear acceleration: communication overhead and synchronisation latencies can degrade efficiency, especially without high‑bandwidth interconnects.

What this does NOT solve

Multi‑GPU setups are not a silver bullet. Idle memory slices, network latency and imbalanced workloads often lead to underutilisation. Without careful partitioning and orchestration, the cost of extra GPUs can outweigh the benefits.

Parallelism Selector

To decide which strategy to use, employ the Parallelism Selector:

  • If model size exceeds single‑GPU memory → choose model parallelism (split layers).
  • If dataset or batch size is large but model fits in memory → choose data parallelism (replicate model).
  • If both model and dataset sizes push limits → adopt pipeline parallelism or a hybrid strategy.

Add an extra decision: Check interconnect. If NVLink or InfiniBand isn’t available, the communication cost may negate benefits; consider mid‑tier GPUs or smaller models instead.

Expert insights

  • Utilisation realities: Training GPT‑4 across 25 000 GPUs achieved only 32–36 % utilisation, underscoring the difficulty of maintaining efficiency at scale.
  • Mid‑tier value: For smaller models, GPUs like A10G or T4 deliver better price–performance than H100s.
  • Research breakthroughs: The PDLP distributed algorithm uses grid partitioning and random shuffling to reduce communication overhead.

Quick summary

Question – When do single GPUs hit a wall, and how do we decide on parallelism?

Summary – Single GPUs run out of memory when model size × VRAM requirement exceeds available capacity. Transformers scale super‑linearly: inference costs 2 × tokens × parameters, while training costs ~6 × parameters per token. Use the Parallelism Selector to choose data, model or pipeline parallelism based on memory and batch size. Beware of underutilisation due to communication overhead.

Single‑GPU vs Multi‑GPU Performance & Efficiency

Context: when one card isn’t enough

In the early stages of product development, a single GPU often suffices. Prototyping, debugging and small model training run with minimal overhead and lower cost. Single‑GPU inference can also meet strict latency budgets for interactive applications because there’s no cross‑device communication. But as models grow and data explodes, single GPUs become bottlenecks.

Multi‑GPU clusters, by contrast, can reduce training time from months to days. For example, training a 175 B parameter model may require splitting layers across dozens of cards. Multi‑GPU setups also improve utilisation—clusters maintain > 80 % utilisation when orchestrated effectively, and they process workloads up to 50× faster than single cards. However, clusters introduce complexity: you need high‑bandwidth interconnects (NVLink, NVSwitch, InfiniBand) and distributed storage and must manage inter‑GPU communication.

Operational considerations: measuring real efficiency

Measuring performance isn’t as simple as counting FLOPs. Evaluate:

  • Throughput per GPU: How many tokens or samples per second does each GPU deliver? If throughput drops as you add GPUs, communication overhead may dominate.
  • Latency: Pipeline parallelism adds latency; small batch sizes may suffer. For interactive services with sub‑300 ms budgets, multi‑GPU inference can struggle. In such cases, smaller models or Clarifai’s local runner can run on-device or on mid‑tier GPUs.
  • Utilisation: Use orchestration tools to monitor occupancy. Clusters that maintain > 80 % utilisation justify their cost; underutilised clusters burn cash.

cost‑performance trade‑offs

High utilisation is the economic lever. Suppose a cluster costs US$8/hr per GPU but reduces training time from six months to two days. If time‑to‑market is critical, the payback is clear. For inference, the picture changes: because inference accounts for 80–90 % of spending, throughput per watt matters more than raw speed. It may be cheaper to serve high volumes on well‑utilised multi‑GPU clusters, but low‑volume workloads benefit from single GPUs or serverless inference.

What this does NOT solve

Don’t assume that doubling GPUs halves your training time. Idle slices and synchronisation overhead can waste capacity. Building large on‑prem clusters without FinOps discipline invites capital misallocation and obsolescence; cards depreciate quickly and generational leaps shorten economic life.

Utilisation Efficiency Curve

Plot GPU count on the x‑axis and utilisation (%) on the y‑axis. The curve rises quickly at first, then plateaus and may even decline as communication costs grow. The optimal point—where incremental GPUs deliver diminishing returns—marks your economically efficient cluster size. Orchestration platforms like Clarifai’s compute orchestration can help you operate near this peak by queueing jobs, dynamically batching requests and shifting workloads between clusters.

Expert insights

  • Idle realities: Single GPUs sit idle 70 % of the time on average; clusters maintain 80 %+ utilisation when properly managed.
  • Time vs money: A single GPU would take decades to train GPT‑3, while distributed clusters cut the timeline to weeks or days.
  • Infrastructure: Distributed systems require compute nodes, high‑bandwidth interconnects, storage and orchestration software.

Quick summary

Question – What are the real performance and efficiency trade‑offs between single‑ and multi‑GPU systems?

Summary – Single GPUs are suitable for prototyping and low‑latency inference. Multi‑GPU clusters accelerate training and improve utilisation but require high‑bandwidth interconnects and careful orchestration. Plotting a utilisation efficiency curve helps identify the economically optimal cluster size.

Cost Economics – Capex vs Opex & Unit Economics

Context: what GPUs really cost

Beyond hardware prices, building AI infrastructure means paying for power, cooling, networking and talent. A single H100 costs US$25 K–US$40 K; eight of them in a server cost US$200 K–US$400 K. Upgrading power distribution can run US$10 K–US$50 K, cooling upgrades US$15 K–US$100 K and operational overhead adds US$2–US$7/hr per GPU. True cluster cost therefore lands around US$8–US$15/hr per GPU. On the renting side, marketplace rates in early 2026 are US$0.18/hr for an RTX 4090 and ~US$0.54/hr for an H100 NVL. Given these figures, buying is only cheaper if you sustain high utilisation.

Operational calculation: cost per token and break‑even points

Unit economics isn’t just about the hardware sticker price; it’s about cost per million tokens. A 7 B parameter model must achieve ~50 % utilisation to beat an API’s cost; a 13 B model needs only 10 % utilisation due to economies of scale. Using Clarifai’s dashboards, teams monitor cost per inference or per thousand tokens and adjust accordingly. The Unit‑Economics Calculator framework works as follows:

  1. Input: GPU rental rate or purchase price, electricity cost, model size, expected utilisation hours.
  2. Compute: Total cost over time, including depreciation (e.g., selling a US$1,200 RTX 4090 for US$600 after two years).
  3. Output: Cost per hour and cost per million tokens. Compare to API costs to determine break‑even.

This granular view reveals counterintuitive results: owning an RTX 4090 makes sense only when average utilisation exceeds 4–6 hours/day. For sporadic workloads, renting wins. For inference at scale, multi‑GPU clusters can deliver low cost per token when utilisation is high.

logic for buy vs rent decisions

The logic flows like this: If your workload runs < 4 hours/day or is bursty → rent. If you need constant compute > 6 hours/day for multiple years and can absorb capex and depreciation → buy. If you need multi‑GPU or high‑VRAM jobs → rent because the capital outlay is prohibitive. If you need a mix → adopt a hybrid model: own a small rig, rent for big spikes. Clarifai’s customers often combine local runners for small jobs with remote orchestration for heavy training.

What this does NOT solve

Buying hardware doesn’t protect you from obsolescence; new GPU generations like H200 or MI400 deliver 4× speedups, shrinking the economic life of older cards. Owning also introduces fixed electricity costs—~US$64 per month per GPU at US$0.16/kWh—regardless of utilisation.

Expert insights

  • Investor expectations: Startups that fail to articulate GPU COGS (cost of goods sold) see valuations 20 % lower. Investors expect margins to improve from 50–60 % to ~82 % by Series A.
  • True cost: A 8×H100 cluster costs US$8–US$15/hr after including operational overhead.
  • Marketplace trends: H100 rental prices dropped from US$8/hr to US$2.85–US$3.50/hr; A100 prices sit at US$0.66–US$0.78/hr.

Quick summary

Question – How do I calculate whether to buy or rent GPUs?

Summary – Factor in the full cost: hardware price, electricity, cooling, networking and depreciation. Owning pays off only above about 4–6 hours of daily utilisation; renting makes sense for bursty or multi‑GPU jobs. Use a unit‑economics calculator to compare cost per million tokens and break‑even points.

Inference vs Training – Where do costs accrue?

Context: inference dominates the bill

It’s easy to obsess over training cost, but in production inference usually dwarfs it. According to the FinOps Foundation, inference accounts for 80–90 % of total AI spend, especially for generative applications serving millions of daily queries. Teams that plan budgets around training cost alone find themselves hemorrhaging money when latency‑sensitive inference workloads run around the clock.

Operational practices: boosting inference efficiency

Clarifai’s experience shows that inference workloads are asynchronous and bursty, making autoscaling tricky. Key techniques to improve efficiency include:

  • Server‑side batching: Combine multiple requests into a single GPU call. Clarifai’s inference API automatically merges requests when possible, increasing throughput.
  • Caching: Store results for repeated prompts or subqueries. This is crucial when similar requests recur.
  • Quantisation and LoRA: Use lower‑precision arithmetic (INT8 or 4‑bit) and low‑rank adaptation to cut memory and compute. Clarifai’s platform integrates these optimisations.
  • Dynamic pooling: Share GPUs across services via queueing and priority scheduling. Dynamic scheduling can raise utilisation from 15–30 % to 60–80 %.
  • FinOps dashboards: Track cost per inference or per thousand tokens, set budgets and trigger alerts. Clarifai’s dashboard helps FinOps teams spot anomalies and adjust budgets on the fly.

linking throughput, latency and cost

The economic logic is straightforward: If your inference traffic is steady and high, invest in batching and caching to reduce GPU invocations. If traffic is sporadic, consider serverless inference or small models on mid‑tier GPUs to avoid paying for idle resources. If latency budgets are tight (e.g., interactive coding assistants), larger models may degrade user experience; choose smaller models or quantised versions. Finally, rightsizing—choosing the smallest model that satisfies quality needs—can reduce inference cost dramatically.

What this does NOT solve

Autoscaling isn’t free. AI workloads have high memory consumption and latency sensitivity; spiky traffic can trigger over‑provisioning and leave GPUs idle. Without careful monitoring, autoscaling can backfire and burn money.

Inference Efficiency Ladder

A simple ladder to climb toward optimal inference economics:

  1. Quantise and prune. If your accuracy drop is acceptable (< 1 %), apply INT8 or 4‑bit quantisation and pruning to shrink models.
  2. LoRA fine‑tuning. Use low‑rank adapters to customise models without full retraining.
  3. Dynamic batching and caching. Merge requests and reuse outputs to boost throughput.
  4. GPU pooling and scheduling. Share GPUs across services to maximise occupancy.

Each rung yields incremental savings; together they can reduce inference costs by 30–40 %.

Expert insights

  • Idle cost: A fintech firm wasted US$15 K–US$40 K per month on idle GPUs due to poorly configured autoscaling. Dynamic pooling cut costs by 30 %.
  • FinOps practices: Cross‑functional governance—engineers, finance and executives—helps monitor unit economics and apply optimisation levers.
  • Inference dominance: Serving millions of queries means inference spending dwarfs training.

Quick summary

Question – Where do AI compute costs really accumulate, and how can inference be optimised?

Summary – Inference typically consumes 80–90 % of AI budgets. Techniques like quantisation, LoRA, batching, caching and dynamic pooling can raise utilisation from 15–30 % to 60–80 %, dramatically reducing costs. Autoscaling alone isn’t enough; FinOps dashboards and rightsizing are essential.

Optimisation Levers – Techniques to tame costs

Context: low‑hanging fruit and advanced tricks

Hardware scarcity means software optimisation matters more than ever. Luckily, innovations in model compression and adaptive scheduling are no longer experimental. Quantisation reduces precision to INT8 or even 4‑bit, pruning removes redundant weights, and Low‑Rank Adaptation (LoRA) allows fine‑tuning large models by learning small adaptation matrices. Combined, these techniques can shrink models by up to 4× and speed up inference by 1.29× to 1.71×.

Operational guidance: applying the levers

  1. Choose the smallest model: Before compressing anything, start with the smallest model that meets your task requirements. Clarifai’s model zoo includes small, medium and large models, and its routing features allow you to call different models per request.
  2. Quantise and prune: Use built‑in quantisation tools to convert weights to INT8/INT4. Prune unnecessary parameters either globally or layer‑wise, then re‑train to recover accuracy. Monitor accuracy impact at each step.
  3. Apply LoRA: Fine‑tune only a subset of parameters, often < 1 % of the model, to adapt to your dataset. This reduces memory and training time while maintaining performance.
  4. Enable dynamic batching and caching: On Clarifai’s inference platform, simply setting a parameter turns on server‑side batching; caching repeated prompts is automatic for many endpoints.
  5. Measure and iterate: After each optimisation, check throughput, latency and accuracy. Cost dashboards should display cost per inference to confirm savings.

trade‑offs and decision logic

Not all optimisations suit every workload. If your application demands exact numerical outputs (e.g., scientific computation), aggressive quantisation may degrade results—skip it. If your model is already small (e.g., 3 B parameters), quantisation might yield limited savings; focus on batching and caching instead. If latency budgets are tight, batching may increase tail latency—compensate by tuning batch sizes.

What this does NOT solve

No amount of optimisation will overcome poorly aligned models. Using the wrong architecture for your task wastes compute even if it’s quantised. Similarly, quantisation and pruning aren’t plug‑and‑play; they can cause accuracy drops if not carefully calibrated.

Cost‑Reduction Checklist

Use this step‑by‑step checklist to ensure you don’t miss any savings:

  1. Model selection: Start with the smallest viable model.
  2. Quantisation: Apply INT8 → check accuracy; apply INT4 if acceptable.
  3. Pruning: Remove unimportant weights and re‑train.
  4. LoRA/PEFT: Fine‑tune with low‑rank adapters.
  5. Batching & caching: Enable server‑side batching; implement KV‑cache compression.
  6. Pooling & scheduling: Pool GPUs across services; set queue priorities.
  7. FinOps dashboard: Monitor cost per inference; adjust policies regularly.

Expert insights

  • Clarifai engineers: Quantisation and LoRA can cut costs by around 40 % without new hardware.
  • Photonic future: Researchers demonstrated photonic chips performing convolution at near‑zero energy consumption; while not mainstream yet, they hint at long‑term cost reductions.
  • N:M sparsity: Combining 4‑bit quantisation with structured sparsity speeds up matrix multiplication by 1.71× and reduces latency by 1.29×.

Quick summary

Question – What optimisation techniques can significantly reduce GPU costs?

Summary – Start with the smallest model, then apply quantisation, pruning, LoRA, batching, caching and scheduling. These levers can cut compute costs by 30–40 %. Use a cost‑reduction checklist to ensure no optimisation is missed. Always measure accuracy and throughput after each step.

Model Selection & Routing – Using smaller models effectively

Context: token count drives cost more than parameters

A hidden truth about LLMs is that context length dominates costs. Doubling from a 32 K to a 128 K context triples the memory required for the key/value cache. Similarly, prompting models to “think step‑by‑step” can generate long chains of thought that chew through tokens. In real‑time workloads, large models struggle to maintain high efficiency because requests are sporadic and cannot be batched. Small models, by contrast, often run on a single GPU or even on device, avoiding the overhead of splitting across multiple cards.

Operational tactics: tiered stack and routing

Adopting a tiered model stack is like using the right tool for the job. Instead of defaulting to the largest model, route each request to the smallest capable model. Clarifai’s model routing allows you to set rules based on task type:

  • Tiny local model: Handles simple classification, extraction and rewriting tasks at the edge.
  • Small cloud model: Manages moderate reasoning with short context.
  • Medium model: Tackles multi‑step reasoning or longer context when small models aren’t enough.
  • Large model: Reserved for complex queries that small models cannot answer. Only a small fraction of requests should reach this tier.

Routing can be powered by a lightweight classifier that predicts which model will succeed. Research shows that such Universal Model Routing can dramatically cut costs while maintaining quality.

why small is powerful

Smaller models deliver faster inference, lower latency and higher utilisation. If latency budget is < 300 ms, a large model might never satisfy user expectations; route to a small model instead. If accuracy difference is marginal (e.g., 2 %), favour the smaller model to save compute. Distillation and Parameter‑Efficient Fine‑Tuning (PEFT) closed much of the quality gap in 2025, so small models can tackle tasks once considered out of reach.

What this does NOT solve

Routing doesn’t eliminate the need for large models. Some tasks, such as open‑ended reasoning or multi‑modal generation, still require frontier‑scale models. Routing also requires maintenance; as new models emerge, you must update the classifier and thresholds.

Use‑the‑Smallest‑Thing‑That‑Works (USTTW)

This framework captures the essence of efficient deployment:

  1. Start tiny: Always try the smallest model first.
  2. Escalate only when needed: Route to a larger model if the small model fails.
  3. Monitor and adjust: Regularly evaluate which tier handles what percentage of traffic and adjust thresholds.
  4. Compress tokens: Encourage users to write succinct prompts and responses. Apply token‑efficient reasoning techniques to reduce output length.

Expert insights

  • Default model problem: Teams that pick one large model early and never revisit it leak substantial costs.
  • Distillation works: Research in 2025 showed that distilling a 405 B model into an 8 B version produced 21 % better accuracy on NLI tasks.
  • On‑device tiers: Models like Phi‑4 mini and GPT‑4o mini run on edge devices, enabling hybrid deployment.

Quick summary

Question – How can routing and small models cut costs without sacrificing quality?

Summary – Token count often drives cost more than parameter count. Adopting a tiered stack and routing requests to the smallest capable model reduces compute and latency. Distillation and PEFT have narrowed the quality gap, making small models viable for many tasks.

Multi‑GPU Training – Parallelism Strategies & Implementation

Context: distributing for capacity and speed

Large‑parameter models and massive datasets demand multi‑GPU training. Data parallelism replicates the model and splits the batch across GPUs; model parallelism splits layers; pipeline parallelism stages operations across devices. Hybrid strategies blend these to handle complex workloads. Without multi‑GPU training, training times become impractically long—one article noted that training GPT‑3 on a single GPU would take decades.

Operational steps: running distributed training

A practical multi‑GPU training workflow looks like this:

  1. Choose parallelism strategy: Use the Parallelism Selector to decide between data, model, pipeline or hybrid parallelism.
  2. Set up environment: Install distributed training libraries (e.g., PyTorch Distributed, DeepSpeed). Ensure high‑bandwidth interconnects (NVLink, InfiniBand) and proper topology mapping. Clarifai’s training orchestration automates some of these steps, abstracting hardware details.
  3. Profile communication overhead: Run small batches to measure all‑reduce latency. Adjust batch sizes and gradient accumulation steps accordingly.
  4. Implement checkpointing: For long jobs, especially on pre‑emptible spot instances, periodically save checkpoints to avoid losing work.
  5. Monitor utilisation: Use Clarifai’s dashboards or other profilers to track utilisation. Balance workloads to prevent stragglers.

weighing the trade‑offs

If your model fits in memory but training time is long, data parallelism gives linear speedups at the expense of memory duplication. If your model doesn’t fit, model or pipeline parallelism becomes mandatory. If both memory and compute are bottlenecks, hybrid strategies deliver the best of both worlds. The choice also depends on interconnect; without NVLink, model parallelism may stall due to slow PCIe transfers.

What this does NOT solve

Parallelism can complicate debugging and increase code complexity. Over‑segmenting models can introduce excessive communication overhead. Multi‑GPU training is also power‑hungry; energy costs add up quickly. When budgets are tight, consider starting with a smaller model or renting bigger single‑GPU cards.

Parallelism Playbook

A comparison table helps decision‑making:

Strategy

Memory usage

Throughput

Latency

Complexity

Use case

Data

High (full model on each GPU)

Near‑linear

Low

Simple

Fits memory; large datasets

Model

Low (split across GPUs)

Moderate

High

Moderate

Model too large for one GPU

Pipeline

Low

High

High

Moderate

Sequential tasks; long models

Hybrid

Moderate

High

Moderate

High

Both memory and compute limits

Expert insights

  • Time savings: Multi‑GPU training can cut months off training schedules and enable models that wouldn’t fit otherwise.
  • Interconnect matter: High‑bandwidth networks (NVLink, NVSwitch) minimise communication overhead.
  • Checkpoints and spot instances: Pre‑emptible GPUs are cheaper but require checkpointing to avoid job loss.

Quick summary

Question – How do I implement multi‑GPU training efficiently?

Summary – Decide on parallelism type based on memory and dataset size. Use distributed training libraries, high‑bandwidth interconnects and checkpointing. Monitor utilisation and avoid over‑partitioning, which can introduce communication bottlenecks.

Deployment Models – Cloud, On‑Premise & Hybrid

Context: choosing where to run

Deployment strategies range from on‑prem clusters (capex heavy) to cloud rentals (opex) to home labs and hybrid setups. A typical home lab with a single RTX 4090 costs around US$2,200 plus US$770/year for electricity; a dual‑GPU build costs ~US$4,000. Cloud platforms rent GPUs by the hour with no upfront cost but charge higher rates for high‑end cards. Hybrid setups mix both: own a workstation for experiments and rent clusters for heavy lifting.

Operational decision tree

Use the Deployment Decision Tree to guide choices:

  • Daily usage < 4 h: Rent. Marketplace GPUs cost US$0.18/hr for RTX 4090 or US$0.54/hr for H100.
  • Daily usage 4–6 h for ≥ 18 months: Buy. The initial investment pays off after two years.
  • Multi‑GPU jobs: Rent or hybrid. Capex for multi‑GPU rigs is high and hardware depreciates quickly.
  • Data sensitive: On‑prem. Compliance requirements or low‑latency needs justify local servers; Clarifai’s local runner makes on‑prem inference easy.
  • Regional diversity & cost arbitrage: Multi‑cloud. Spread workloads across regions and providers to avoid lock‑in and exploit price differences; Clarifai’s orchestration layer abstracts provider differences and schedules jobs across clusters.

balancing flexibility and capital

If you experiment often and need different hardware types, renting provides agility; you can spin up an 80 GB GPU for a day and return to smaller cards tomorrow. If your product requires 24/7 inference and data can’t leave your network, owning hardware or using a local runner reduces opex and mitigates data‑sovereignty concerns. If you value both flexibility and baseline capacity, adopt hybrid: own one card, rent the rest.

What this does NOT solve

Deploying on‑prem doesn’t immunise you from supply shocks; you still need to maintain hardware, handle power and cooling, and upgrade when generational leaps arrive. Renting isn’t always available either; spot instances can sell out during demand spikes, leaving you without capacity.

Expert insights

  • Energy cost: Running a home‑lab GPU 24/7 at US$0.16/kWh costs ~US$64/month, rising to US$120/month in high‑cost regions.
  • Hybrid in practice: Many practitioners own one GPU for experiments but rent clusters for large training; this approach keeps fixed costs low and offers flexibility.
  • Clarifai tooling: The platform’s local runner supports on‑prem inference; its compute orchestration schedules jobs across clouds and on‑prem clusters.

Quick summary

Question – Should you deploy on‑prem, in the cloud or hybrid?

Summary – The choice depends on utilisation, capital and data sensitivity. Rent GPUs for bursty or multi‑GPU workloads, buy single cards when utilisation is high and long‑term, and use hybrid when you need both flexibility and baseline capacity. Clarifai’s orchestration layer abstracts multi‑cloud differences and supports on‑prem inference.

Sustainability & Environmental Considerations

Context: the unseen footprint

AI isn’t just expensive; it’s energy‑hungry. Analysts estimate that AI inference could consume 165–326 TWh of electricity annually by 2028—equivalent to powering about 22 % of U.S. households. Training a single large model can use over 1,000 MWh of energy, and generating 1,000 images emits carbon equivalent to driving four miles. GPUs rely on rare earth elements and heavy metals, and training GPT‑4 could consume up to seven tons of toxic materials.

Operational practices: eco‑efficiency

Environmental and financial efficiencies are intertwined. If you raise utilisation from 20 % to 60 %, you can reduce GPU needs by 93 %—saving money and carbon simultaneously. Adopt these practices:

  • Quantisation and pruning: Smaller models require less power and memory.
  • LoRA and PEFT: Update only a fraction of parameters to reduce training time and energy.
  • Utilisation monitoring: Use orchestration to keep GPUs busy; Clarifai’s scheduler offloads idle capacity automatically.
  • Renewable co‑location: Place data centres near renewable energy sources and implement advanced cooling (liquid immersion or AI‑driven temperature optimisation).
  • Recycling and longevity: Extend GPU lifespan through high utilisation; delaying upgrades reduces rare‑material waste.

cost meets carbon

Your power bill and your carbon bill often scale together. If you ignore utilisation, you waste both money and energy. If you can run a smaller quantised model on a T4 GPU instead of an H100, you save on electricity and prolong hardware life. Efficiency improvements also reduce cooling needs; smaller clusters generate less heat.

What this does NOT solve

Eco‑efficiency strategies don’t remove the material footprint entirely. Rare earth mining and chip fabrication remain resource‑intensive. Without broad industry change—recycling programs, alternative materials and photonic chips—AI’s environmental impact will continue to grow.

Eco‑Efficiency Scorecard

Rate each deployment option across utilisation (%), model size, hardware type and energy consumption. For example, a quantised small model on a mid‑tier GPU with 80 % utilisation scores high on eco‑efficiency; a large model on an underutilised H100 scores poorly. Use the scorecard to balance performance, cost and sustainability.

Expert insights

  • Energy researchers: AI inference could strain national grids; some providers are even exploring nuclear power.
  • Materials scientists: Extending GPU life from one to three years and increasing utilisation from 20 % to 60 % can reduce GPU needs by 93 %.
  • Clarifai’s stance: Quantisation and layer offloading reduce energy per inference and allow deployment on smaller hardware.

Quick summary

Question – How do GPU scaling choices impact sustainability?

Summary – AI workloads consume enormous energy and rely on scarce materials. Raising utilisation and employing model optimisation techniques reduce both cost and carbon. Co‑locating with renewable energy and using advanced cooling further improve eco‑efficiency.

Emerging Hardware & Alternative Compute Paradigms

Context: beyond the GPU

While GPUs dominate today, the future is heterogeneous. Mid‑tier GPUs handle many workloads at a fraction of the cost; domain‑specific accelerators like TPUs, FPGAs and custom ASICs offer efficiency gains; AMD’s MI300X and upcoming MI400 deliver competitive price–performance; photonic or optical chips promise 10–100× energy efficiency. Meanwhile, decentralised physical infrastructure networks (DePIN) pool GPUs across the globe, offering cost savings of 50–80 %.

Operational guidance: evaluating alternatives

  • Match hardware to workload: Matrix multiplications benefit from GPUs; convolutional tasks may run better on FPGAs; search queries can leverage TPUs. Clarifai’s hardware‑abstraction layer helps deploy models across GPUs, TPUs or FPGAs without rewriting code.
  • Assess ecosystem maturity: TPUs and FPGAs have smaller developer ecosystems than GPUs. Ensure your frameworks support the hardware.
  • Consider integration costs: Porting code to a new accelerator may require engineering effort; weigh this against potential savings.
  • Explore DePIN: If your workload is tolerant of variable latency and you can encrypt data, DePIN networks provide massive capacity at lower prices—but evaluate privacy and compliance risks.

When to adopt

If GPU supply is constrained or too expensive, exploring alternative hardware makes sense. If your workload is stable and high volume, porting to a TPU or custom ASIC may offer long‑term savings. If you need elasticity and low commitment, DePIN or multi‑cloud strategies let you arbitrage pricing and capacity. But early adoption can suffer from immature tooling; consider waiting until software stacks mature.

What this does NOT solve

Alternative hardware doesn’t fix fragmentation. Each accelerator has its own compilers, toolchains and limitations. DePIN networks raise latency and data‑privacy concerns; secure scheduling and encryption are essential. Photonic chips are promising but not yet production‑ready.

Hardware Selection Radar

Visualise accelerators on a radar chart with axes for cost, performance, energy efficiency and ecosystem maturity. GPUs score high on maturity and performance but medium on cost and energy. TPUs score high on efficiency and cost but lower on maturity. Photonic chips show high potential on efficiency but low current maturity. Use this radar to identify which accelerator aligns with your priorities.

Expert insights

  • Clarifai roadmap: The platform will integrate photonic and alternative accelerators, abstracting complexity for developers.
  • DePIN projections: Decentralised GPU networks could generate US$3.5 T by 2028; 89 % of organisations already use multi‑cloud strategies.
  • XPUs rising: Enterprise spending on TPUs, FPGAs and ASICs is growing 22.1 % YoY.

Quick summary

Question – When should AI teams consider alternative hardware or DePIN?

Summary – Explore alternative accelerators when GPUs are scarce or costly. Match workloads to hardware, evaluate ecosystem maturity and integration costs, and consider DePIN for price arbitrage. Photonic chips and MI400 promise future efficiency but are still maturing.

Conclusion & Recommendations

Synthesising the journey

The economics of AI compute are shaped by scarcity, super‑linear scaling and hidden costs. GPUs are expensive not only because of high‑bandwidth memory constraints but also due to lead times and vendor prioritisation. Single GPUs are perfect for experimentation and low‑latency inference; multi‑GPU clusters unlock large models and faster training but require careful orchestration. True cost includes power, cooling and depreciation; owning hardware makes sense only above 4–6 hours of daily use. Most spending goes to inference, so optimising quantisation, batching and routing is paramount. Sustainable computing demands high utilisation, model compression and renewable energy.

Recommendations: the Scale‑Right Decision Tree

Our final framework synthesises the article’s insights into a practical tool:

  1. Assess demand: Estimate model size, context length and daily compute hours. Use the GPU Economics Stack to identify demand drivers (tokens, parameters, context).
  2. Check supply and budget: Evaluate current GPU prices, availability and lead times. Decide if you can secure cards or need to rent.
  3. Right‑size models: Apply the Use‑the‑Smallest‑Thing‑That‑Works framework: start with small models, use routing to call larger models only when necessary.
  4. Decide on hardware: Use the Capex vs Opex Decision Matrix and Hardware Selection Radar to choose between on‑prem, cloud or hybrid and evaluate alternative accelerators.
  5. Choose parallelism strategy: Apply the Parallelism Selector and Parallelism Playbook to pick data, model, pipeline or hybrid parallelism.
  6. Optimise execution: Run through the Cost‑Reduction Checklist—quantise, prune, LoRA, batch, cache, pool, monitor—keeping the Inference Efficiency Ladder in mind.
  7. Monitor and iterate: Use FinOps dashboards to track unit economics. Adjust budgets, thresholds and routing as workloads evolve.
  8. Consider sustainability: Evaluate your deployment using the Eco‑Efficiency Scorecard and co‑locate with renewable energy where possible.
  9. Stay future‑proof: Watch the rise of DePIN, TPUs, FPGAs and photonic chips. Be ready to migrate when they deliver compelling cost or energy benefits.

Final thoughts

Compute is the oxygen of AI, but oxygen isn’t free. Winning in the AI arms race means more than buying GPUs; it requires strategic planning, efficient algorithms, disciplined financial governance and a willingness to embrace new paradigms. Clarifai’s platform embodies these principles: its compute orchestration pools GPUs across clouds and on‑prem clusters, its inference API dynamically batches and caches, and its local runner brings models to the edge. By combining these tools with the frameworks in this guide, your organisation can scale right—delivering transformative AI without suffocating under hardware costs.

 



Love Eternal’s Uncanny Mix of Precision Platformers and Psychological Horror


Summary

  • Run, jump and flip gravity to escape the mind of a selfish god.
  • Themes of isolation and childhood struggles in a treacherous world.
  • Experimental narrative channeling influences from Silent Hill to Satoshi Kon’s filmography.

Love Eternal, releasing on Xbox One today,is a psychological horror platforming game set in a mysterious and ancient dilapidated castle, which serves as the prison of our abductee protagonist Maya. The predicament of Maya, an American teenager who gets spirited away in the middle of dinner from her suburban family home by a callous and jealous god, is quite the emotional whiplash for anyone, but to what end must Maya endure these trials?

Today I’ll dig into the diverse influences and ingredients that the developers of Love Eternal, brlka, mixed together in order to serve up an unsettling journey with an experimental narrative where you truly won’t know what next to expect.

For starters, brlka, comprised of siblings Toby and Sam Alden, conceived of Love Eternal as an evolution to an older platforming game they made together, simply titled Love. The original Love, made almost a decade ago, shares mechanical roots with Love Eternal in that both are platformers in the vein of masocore (a portmanteau of masochism and hardcore) inspired from the 2000s freeware era of games like Jumper, an early platforming game by Maddy Thorson long before the likes of Celeste, and the Knytt series from Swedish developer Nifflas.

These masocore challenges – and the effort required to overcome your initial apprehensions at seemingly insurmountable platforming arcs – traditionally serve as a meditation on a rhythmic cycle of failing, observing one’s mistakes and mental lapses, and returning to the breach. With Love Eternal, however, brutality in platforming is the scaffolding for Toby, the game’s principal programmer and designer, to hang the quiet terror of the game’s narrative dressing.

Love Eternal’s influences gel together with considered curation – yet it’s also accurate to say, “These are just the movies and TV that Toby was watching at the time of development!” Either way you see it, the influences that run in Love Eternal’s blood include significant works from the esoteric works of Kiyoshi Kurosawa, particularly “Pulse”  (2001); the fever dreams of “Twin Peaks” season 3 and “Paprika” (alongside all of Satoshi Kon’s other animated movies); and the inimitable 1998 anime “Serial Experiments Lain”.

The common thread is how these pieces of media parlay psychological horror, even though horror wouldn’t be considered each work’s principal genre. As it appeared to Toby, psychological horror was seemingly underrepresented in precision platforming, and they decided to use horror to express the narrative in deeply uncanny ways.

Toby has described the essence of the Love Eternal experience as the feeling that one gets after experiencing a nightmare that, when described to another person, that other person would not find immediately apparent as to what was scary about the nightmare, even though one’s direct participatory experience of the nightmare is charged with a curdling unease.

To my eye, having seen the Aldens repeatedly discuss their development journey, the themes of Love Eternal have emerged organically during development, rather than being explicitly designed into the game. For example, the theme of isolation stands out through the visual design and structural composition, with the cavernous environments of the god’s castle imposing upon the miniscule Maya who is comparatively only a few pixels tall.

The juxtaposition of Maya – who is still a child – and her isolation in the vast castle ruins viscerally convey how children often lack physical and mental autonomy, and must submit to the whims and mercies of higher powers, parents or otherwise. Taken from the view of an adult – and likely worsened if a player has the experience of parenthood – Love Eternal’s lens of childhood is abstractly horrific.

There’s horror even in the experience of a child learning to inhabit and occupy their own body as it grows and contorts in unexpected ways, though that horror isn’t explicitly violent. Even simple experiences like losing your baby teeth as new ones push out, or the awkwardness of reconciling with the image of one’s rapidly morphing body during puberty can be strange and unnerving markers of growing up.

Such is the flavor of unusual, lightly grotesque expression of body horror conveyed by Love Eternal’s gorgeously expressive pixel aesthetic and animation. This is all crafted by Sam, a professional animator and artist whose experience spans across work like the aforementioned Love to well-known franchises like “Adventure Time”.

Through character animation, Sam pushes how characters’ bodies can be depicted in pixel art, forming a sense of the uncanny with his subtle explorations of how even mundane things might move and writhe. As an example, not too far into Love Eternal’s introductory stages, Maya will come face to face with a distorted version of her father with freakishly elongated limbs, and even in the confines of pixel art, watching this man creep around like a spider (and not like your friendly neighborhood Spider-Man) tended to raise some hairs amongst those who’ve already experienced the demo of Love Eternal.

This uncanny aesthetic dovetails into Love Eternal’s mechanical and narrative pillars for  an experience that careens from somber, melancholic and ominous to bizarre, spine-chilling and occasionally unhinged. Ultimately, the team at brlka and I are confident that just about no one will be able to predict the ways in which the experience of Love Eternal unfurls its twisted tendrils.

We’re excited for players to gear up for the unrelenting challenges and eerie narrative of Love Eternal, now that the game’s out on Xbox One today!


LOVE ETERNAL

Ysbryd Games

☆☆☆☆☆

★★★★★


$9.99

$8.49

Wander a castle built of bitter memories in LOVE ETERNAL, a psychological horror platformer with devious trials and an unsettling, experimental narrative.

Run, jump, and reverse the flow of gravity itself to escape the mind of a selfish god in this challenging precision platformer. Play as Maya, a child stolen from her family on the whim of a lonely, forsaken deity, and make your way through over 100 screens filled with spikes, lasers, switches, and traps as you unravel the horrifying secret of your new prison.

Will you find your way home, or wander these halls carved of memory forever?