They Say AI Is for the Rich. Here's the Way Out.

0:00 A $200 AI plan costs 2,400 a year. Cheap for whom? Buying your way out can cost thousands. We'll price both, including the cheaper local escape. We'll follow your money through subscriptions and

0:13 hardware, then the memory trick that changes the shopping list. Start with a price tag that barely changes, crossing borders where people's budgets change enormously. A country's output doesn't

0:24 tell us your wage. It does put that dollar price into perspective. Claude Pro costs $20 monthly. The largest max tier costs 200 before applicable taxes. That's the web price. You get more usage

0:37 at higher tiers and paid plans still have limits. To compare the scale, divide a country's yearly economic output by its population. That's GDP per capita. The World Bank's 2025 world

0:50 figure is around $14,400 per person. For the same year, America is around 90,000. India is around 2,700. Compare 12 payments with those figures. They approach India's entire annual

1:04 output per person. In America, they're under 3%. If a dollar price sounds affordable, check where the buyer lives. Your checkout page crosses borders more easily than your budget. So, that's the

1:16 price comparison. The other countries on screen use the same measure here. Those nominal dollars move with exchange rates and averages hide who gets the money. You can't turn this chart into a count

1:28 of people who can afford AI. I'd use it to challenge a price pitched as small everywhere. Then check actual spending. Only 1% pays for AI. 1% of whom? Bank of America reported payments from about 3%

1:42 of its households in March. You may be paying through channels this data set can't see. The remaining households aren't proven free users. Goodwater's US survey found 46% of AI users paid for at

1:54 least one subscription. Its respondents and method differ from the bank's records, you shouldn't average the two. Check who was counted before repeating a percentage. These results don't

2:04 establish a worldwide 1% conversion rate. This week's 1% figure describes businesses. Ramp reports that its biggest 1% of customers accounts for 80% of enterprise spending at OpenAI and

2:17 Anthropic. That's Ramp's payment sample. It measures where business money concentrates, not how many ordinary people pay. So far, we've separated the price from

2:27 the people paying it. You can see the incentive. A company with repeated workflows can buy far more usage than one person chatting. You can still buy consumer plans, but if you want control

2:38 over the bill, ownership sounds tempting. Say you buy a $4,000 computer to replace a $20 monthly plan. Divide the purchase by the bill you avoid, 200 months before electricity or repairs.

2:51 This is hypothetical with no promise of equal performance. I'd buy for control before claiming claiming savings from that calculation. Your purchase also meets today's market. In its September

3:02 1st check, Tom's Hardware found a 32 GB RTX 5090 around $5,000. Its launch price was just under 2,000. One card doesn't prove every computer tripled, but a build budget based on launch prices can

3:16 fail at checkout. Apple gives you a different route. The processor and graphics share memory. It's Mac Studio with the M5 Ultra starts at $5,499. Apple advertises up to 512 GB, but that

3:32 option is due in late October. The starting price and maximum memory don't describe one verified configuration. You need the quote for the machine you're actually ordering. That leaves the

3:42 workstation ceiling. AMD's Halo Station supports up to 576 GB of graphics memory. It's a prototype coming in 2027. Tom's Hardware estimates a parts bill

3:55 above $100,000, potentially beyond 150,000. You can't treat that estimate as a retail quote. Exact lists a configuration of Nvidia's DGX Station at

4:06 $94,930. You're buying fast graphics memory connected to larger processor memory. Those parts move data at different speeds. Adding the capacities doesn't

4:17 make the slower memory faster. Leave the desk and Mercatus $450,000 for a server with eight B200 GPUs. It's a July estimate repeated in September, including the surrounding machine. Is

4:31 this the capacity your work needs or the capacity a launch photo makes you want? You can rent to remove the upfront hardware bill. Hyperstack lists an H200 at $3.99 per GPU hour. First, choose the

4:44 machine and read what the rate includes. Then count the hours it stays allocated. 10 hours comes to $39.90. You've bought 10 hours of compute, not unlimited AI for the month. Leave that

4:57 rental running for 30 days and you've bought 720 hours. Multiply by the same rate, roughly $2873 before extras. Finally, include storage and the other build resources in your

5:10 budget. The clock explains why a cheap trial can become expensive continuous use. A cloud GPU and a chat subscription are different products. So, that's the cost of capacity.

5:22 Your next question is how much the model needs. Think of graphics memory as a workbench. Step one, choose weights you're allowed to download. Those are the model's learned numerical settings.

5:33 A nominal 27 billion parameter model at 16 bits needs about 54 GB for those settings alone. A download link doesn't supply the bench. Step two, load those settings into the available memory. Step

5:46 three, leave room for the conversation's working notes and the runtime scratch space. The notes include the attention cache, which is the intermediate work saved from tokens already read. Longer

5:57 context can need more space. If your settings fill the bench, there's nowhere to do the work. Your file size alone can't tell you the full requirement. Step four, check what happens when the

6:08 workload exceeds fast memory. Software can move some work to system memory, but that changes the path the data travels. You might accept a slower answer. You might need a shorter conversation.

6:20 More installed RAM helps capacity, but it doesn't turn into graphics memory with the same speed. That's why a bare 24 GB minimum is too crude. So, that's the memory budget. Compression shrinks

6:32 the settings themselves. Imagine marking a position on a ruler with fewer allowed notches. That's quantization, which means storing a coarser value using fewer bits. Four bits would make that

6:44 nominal model 13 and 1/2 GB before overhead. But, careless rounding changes its answers. You need to decide which errors it can tolerate. The GSQ method first compares

6:55 compressed outputs with the original on calibration examples. It learns soft choices between allowed values, then sharpens those choices into stored values. RCO decides which groups receive

7:07 which precision under a fixed-size budget. More bits can go where errors hurt most. The constraint forces the method to spend less elsewhere instead of letting

7:17 the file grow. Here's the file, quant 3.8 27B, compressed with GSQ and RCO. Its smallest standard release is 8.42 GB. You can read that size in the file list. The settings use a smaller

7:32 representation. The model hasn't lost most of its parameters. Follow that file into your machine, and the earlier workbench rules still applies. Load the weights, reserve working memory, then

7:44 run your actual prompt. An 8.42 GB file isn't a promise that an 8 GB graphics card can run everything. The repository also has larger variants and optional prediction heads. Check the exact file

7:58 name before carrying its number into a shopping list. That brings us to a reported run below 24 GB. Velostek reports running an IQ2 extra small variant on a 12 GB Radeon laptop GPU.

8:11 The reported speed was around 21 tokens a second with specific software settings. That's one tester's result, not our test or a guarantee for your laptop. It gives you a configuration to

8:23 investigate before spending thousands. The quality question survives the memory win. A smaller file can lose accuracy, and fitting a longer conversation doesn't prove it remembers that

8:33 conversation well. Before buying hardware, try a task you can score. Your own code with tests, or a document with answers you already know. Check the output and the waiting time together. A

8:45 model that fits but fails your work hasn't saved you money. You now have a purchasing rule. Start with the computer you own and a model small enough to leave working

8:55 room. You can rent briefly to test a larger workload. Buy capacity when repeated use or local control justifies it. If your agent sends files to outside services, owning its model doesn't make

9:07 the whole workflow private. Follow the data as carefully as the bill. My verdict, buy the smallest amount of AI that reliably does your work. For a price-sensitive individual, free access

9:19 and a capable local model on existing hardware win the starting decision. A subscription earns its place when it saves you enough time. Dedicated hardware earns its place

9:29 through sustained use or control. The premium market rewards deep pockets, but you don't have to to its ceiling to use AI. I'd use a smaller local model for work it can handle and rent extra

9:41 capability when needed. Would you rather own a model you can keep using or pay for a stronger service someone else controls?