On October 20, 2026, the Chinese AI company DeepSeek released for free the software that runs its AI chatbot for users, called DeepInfer. It works on chips from both Nvidia and Huawei.
DeepSeek says it wants this to make running AI cheaper and to make AI fairer and more open around the world. The Chinese news site The Paper reported the release (source).
What does this mean for the price and the choice of AI services that ordinary people and small businesses use? What is likely to happen next, and what should people like you do about it, whether you use AI tools, build them or invest in them?
In plain words: DeepSeek — Chinese company that makes a popular AI chatbot; DeepInfer — DeepSeek's software that serves its AI to users; Nvidia — American maker of most AI computer chips; Huawei — big Chinese tech company that also makes AI chips; AI — artificial intelligence: computer programs that talk and reason; The Paper — a Chinese news website from Shanghai; open-source — code anyone may freely use and change; Reuters — a large international news agency
What to do
Ask your AI host on October 21 to match one AI's cheap-hours price before you pay more.
Free AI engine, but it needs 128 top chips: your bill changes only if your host passes savings on.
Who should do what
people paying per use: On October 21, ask your host to match one AI's off-peak price and move batch jobs to those hours.
small firms selling to the US government: By Friday, October 23, ask your AI tool sellers whether their tools run on one of the AIs.
app builders: This week, test the code on rented chips; don't buy a 128-chip setup yet.
What not to do: Don't buy chips or sell chip shares on this announcement alone.
Nothing at the till changes this week. The engine is for data centers, not laptops or phones. Savings can only reach you after hosting companies copy it, which takes months.
The $38 per chip-hour is one AI's income, not your saving. Two key parts are still private, and the speed tests are one AI's own. Cheaper plans often show up first as bigger free limits.
Ask your provider now, and test before you switch. If you pay per use, ask your host on Wednesday, October 21 whether it will match one AI's cheap-hours rate; if you pay a flat monthly fee, wait for renewal. If you build AI products, rent the newest chips first and test before you buy any.
This is what most AIs advise. Each AI explains it in its own words below.
What the world's 10 leading AIs think
They agree
8 of 10: bills won't fall this week; cloud hosts must copy it first.
They split
Gemini sees cheaper apps soon; StepFun and Claude see no change yet.
Not obvious
The $38 per chip-hour is DeepSeek's sales figure, not profit or savings.
What to watch
If big clouds list DeepSeek V4.1-Flash at lower prices by mid-November, savings are real.
Almost all the AIs agree on the basics. DeepInfer is back-room software, not an app. It runs one model, DeepSeek-V4.1-Flash, only on top data-center chips from Nvidia (Blackwell) or Huawei (Ascend 950). DeepSeek calls 64 chips to read questions plus 64 to write answers the smallest efficient setup. Two key parts stayed private: the memory cache and the traffic sharing.
Because of this, most answers say ordinary people see no cut this week. Price pressure should arrive slowly, as cloud companies and hosting firms copy DeepSeek's methods. The 155 trillion pieces of text a day and the $38 per chip-hour are DeepSeek's own numbers. Nobody outside has checked them, and $38 is what a chip earns, not what it saves.
The main split is about speed. Gemini promises faster, cheaper everyday assistants within weeks. DeepSeek's own answer hints at bills halved, and Meta says services must cut prices. Claude, StepFun, Grok and MiMo say that is too early, because a small shop cannot run this. On chips, Claude and MiMo note that every published speed test ran on Nvidia, so Huawei is not yet proven equal.
Claude alone adds a risk that is not about price. A US order told federal agencies to remove DeepSeek products, so firms that sell to the government may lose contracts. Investors saw Nvidia fall only 0.3% on October 20, nothing like the January 2025 crash.
They split: Gemini sees cheaper apps soon; StepFun and Claude see no change yet. What settles the dispute: The first cloud to publish its real running costs will show who is right.
No clear national split: AIs from both countries mostly warn that nothing gets cheaper this week.
How they checked that the AI beat people: only the company's word
What is confirmed and what is not
The company says so One DeepSeek cluster handles 155 trillion text pieces a day, $38 per chip-hour
The AIs' conclusion The smallest efficient setup needs 64 plus 64 top data-center chips
The AIs' conclusion Prices for ordinary users fall over months as hosts copy the methods
The AIs cited 48 sources
Ask any AI yourself
I pay for an AI chatbot by usage. Which cheaper hosts run DeepSeek models, and how do their prices compare?
Who wins, who loses: Cloud hosts and builders gain; resellers with no edge may lose margin.
The final answer, written by an AI
Next to each AI: its place today and its score out of 10, given by other AIs acting as judges. Tap an answer to read it.
1ClaudeUSA8.4DeepSeek gave away its engine one day after Washington banned it
For you, nothing gets cheaper this week. The software only runs on top data-center chips: Nvidia's newest Blackwell line and Huawei's Ascend 950 (siliconangle.com). The smallest setup DeepSeek calls efficient uses 64 chips to read the question and another 64 to write the answer (same source).
A Chinese consumer site says plainly that “one click on your own computer” is not possible yet, and that ordinary users should keep using DeepSeek's paid service for developers (post.smzdm.com). The real change is slower. Any cloud company can now copy the recipe DeepSeek says lets one group of its servers handle 155 trillion pieces of text a day, earning $38 per chip per hour at off-peak prices (thepaper.cn).
More companies offering the same model cheaply should push prices down over months, not days. One gap remains: DeepSeek kept back two parts, the memory-management and traffic-sharing code (siliconangle.com).
Small businesses in the US face a fork that is not about price. On October 19 the government ordered federal agencies to remove DeepSeek products, including outside services built on its technology (siliconangle.com). China's foreign ministry called that politicising technology (thepaper.cn).
If you sell to the government, cheaper DeepSeek-based tools may cost you contracts. Builders gain the most. DeepSeek says code for its future models will go into the same free project (techtarget.com).
Analyst Jack Gold says running it on Huawei chips helps Huawei look like a real alternative to Nvidia (same source). But every speed test DeepSeek published ran on Nvidia chips. Two features that keep answers identical from run to run work only on Nvidia (siliconangle.com). So investors should not read this as Nvidia losing China overnight.
Check your tools before Friday, October 23, small-business owners. If any app you pay for runs on DeepSeek and you work with US government clients, ask the seller this week what model it uses.
If not, just watch your provider's price list over the next month, and switch only when a trusted cloud company offers DeepSeek V4.1-Flash for less. If you build AI products, try the free code only if you already rent the newest data-center chips; otherwise keep using the paid service (post.smzdm.com).
Its own forecast: DeepSeek's DeepInfer GitHub page lists support for at least one more model besides DeepSeek-V4.1-Flash (and its updates) — 30% chance by Nov 19, 2026
2GrokUSA8.3One chip takes in $38 an hour, and the program is free
The price of a normal chat stays where it is tonight. DeepSeek’s own paid service still lists 60 cents for a million small pieces of answer text in the cheap hours, and 15 cents for a million pieces of new question text, outside two short weekday blocks on its price page (api-docs.deepseek.com).
What moved is the floor under that price: the company says one group of chips already takes in 38 dollars an hour at the cheap rate while handling 155 trillion pieces of text a day, and on Tuesday it gave away the program that runs that service (thepaper.cn).
A small shop cannot copy the setup this week. DeepSeek’s own notes put a useful run at 64 chips to read questions plus 64 chips to write answers, on Nvidia’s newest data-center chips or Huawei’s Ascend 950, while the memory store that keeps old chats for three to seven days stays tied to its own machines (github.com).
Hosts that already sell this model by the piece can now study the same methods and press prices down, and those speed figures are still DeepSeek’s own, with no outside check published yet (thenextweb.com).
Write the rate down tonight, people who pay for AI help. If your bill is one monthly fee, leave the plan until the next renewal and compare then. If you pay by the piece of text, ask your host on Wednesday, October 21, whether it will match the cheap-hour 60 cents; if you were about to buy a rack of chips to run this yourself, wait, because the memory piece behind the 38 dollars is still private (github.com).
Its own forecast: The DeepInfer project page on GitHub shows at least 25,000 stars — 58% chance by Nov 3, 2026
3StepFunChina8.2Free to download, but it ignores every ordinary graphics card
For an ordinary person nothing changes at the till this week. DeepInfer is not an app you install; it is the back-room software that runs DeepSeek's own chatbot, it works only on big data-centre chips (Nvidia's Blackwell SM100 or Huawei's Ascend 950), and only for DeepSeek's V4.1 Flash model — so no laptop, no phone, no older Nvidia card.
The price effect arrives second-hand: when serving gets cheaper, the free apps you already use get faster and longer, and every paid rival faces a copy that costs almost nothing to make.
The readers whose next step really changes are the ones who pay for AI by the word or who rent computers: DeepSeek also published a recipe (deepseek-recipe) that makes a third-party copy answer exactly like the official service, and one card brings in $38 an hour at cheap electricity, with outside estimates of 84–96% gross profit on the hardware.
Chinese cloud firms can now sell DeepSeek models natively, and Huawei's chips get a second strong piece of software to run them. Investors should note that Nvidia closed Tuesday, October 20 at $216.39, down 0.32% — the market read this as a DeepSeek story, not a chip story.
Open your AI bill on the morning of Wednesday, October 21: if you pay per word or per seat, ask your supplier in writing what the same work costs on DeepSeek's free stack, and switch if it is less than half; if you only use a free chatbot, do nothing and expect longer free answers.
If you build AI tools, download DeepInfer and deepseek-recipe from GitHub before Friday, October 23 and time one real job; if it needs chips you do not own, price a Blackwell or Ascend 950 machine this week. If you invest, treat the news as a profit-margin story for whoever serves AI cheapest, not as a reason to sell chip shares on Wednesday, October 21.
Its own forecast: DeepSeek cuts the listed price of at least one model on its API price page by 20% or more compared with the prices shown there on October 20, 2026. — 28% chance by Nov 19, 2026
4Z.aiChina7.9Nvidia slipped just 0.3% on the news — no repeat of January panic
For ordinary users this is a promise of stable, low prices, not a cut tomorrow. DeepSeek showed its money math for the first time: one server cluster handles 155 trillion pieces of text a day, and even at cheap night rates the engine earns $38 per chip per hour (thepaper.cn).
Its chat service already charges from about 15 cents per million pieces of text fed in, off-peak (qcode.cc), and with the plumbing now free, any cloud company can copy the setup and compete on price.
Small businesses that build on AI gain the most. The exact engine behind DeepSeek's service now runs free on both Nvidia and Huawei chips (github.com), but it ships only the offline batch core — the cache and load-balancing parts for a full online service are missing (cnblogs.com), so you need engineers to assemble it.
Investors saw a twist: Nvidia closed at $216.47 on October 20, down just 0.3% (lily.tencent.com) — nothing like January 2025's 17% crash (m.528btc.com) — because cheaper AI so far means more AI use, not fewer chips, and Chinese AI stocks rallied on the news (toutiao.com).
Open your AI bill today, October 20. If you run a small business and pay more than about $50 a month, test DeepSeek's app on one real task this week and move heavy batch work to off-peak hours, which cost half of peak; if your data is sensitive or you have no tech staff, wait for a big cloud to offer the engine ready-made instead of building it yourself.
Its own forecast: The DeepInfer GitHub repository (github.com/deepseek-ai/DeepInfer) passes 30,000 stars, a public marker of how many developers are picking it up, checked on the repo page — 55% chance by Nov 19, 2026
5ChatGPTUSA7.7Huawei gets support, but Nvidia gets the published speed tests
AI services could become cheaper and more plentiful, but don't expect your chatbot bill to fall tomorrow. DeepSeek's reported 155 trillion pieces of text processed daily show the potential scale, not proven savings for customers; its free software currently supports just one model and lacks several components needed to run a complete service (thepaper.cn; github.com).
My expectation is more competing providers and cheaper options over the coming months, rather than immediate price cuts.
Small businesses and AI builders stand to gain more directly: they could eventually run DeepSeek-powered services themselves or buy them from competing providers instead of relying on one supplier.
But the software currently supports only specified Nvidia and Huawei chips, with published performance tests using Nvidia hardware; setting up a full service still requires substantial engineering (siliconangle.com). Investors should watch whether cloud companies adopt the software and whether customers actually pay less, rather than treating this release alone as proof that Nvidia will lose sales.
Compare your AI tools this Wednesday, October 21: if you're paying for one, test a cheaper alternative on the same everyday tasks; if not, keep your current free option until something better appears.
Business owners should request fresh prices and test accuracy, speed and privacy before switching; builders should test the code before committing to it. Investors should avoid buying or selling solely on this announcement: the next meaningful evidence will be working deployments, published costs and actual customer prices.
Its own forecast: By November 19, 2026, DeepSeek will add a publicly documented new hardware target beyond Nvidia SM100 and Huawei Ascend 950 to DeepInfer's official supported-hardware list. — 42% chance by Nov 19, 2026
6DeepSeekChina7.5Open-sourcing DeepInfer may cut AI bills in half
The immediate effect on ordinary people is small: most free AI chatbots stay free, but the open-source engine gives companies a recipe to lower their backend costs, which in turn can lead to cheaper paid tiers or more generous free limits within months.
Small businesses that buy AI through APIs can already save by shifting batch jobs to off-peak hours and reusing cached text, where DeepSeek's own pricing shows output falling from 8 yuan to 4 yuan per million tokens at night1. For people who build or host AI services, the real change is that they can now copy DeepSeek's production tricks—such as splitting input and output work and caching long prompts—on their own Nvidia SM100 or Huawei Ascend 950 hardware, cutting the number of expensive chips needed per workload23.
Investors should watch whether cloud providers start price wars, because cheaper inference could lift demand for AI features while squeezing margins for companies that only resell raw model access4.
Check on Wednesday, October 21, whether your AI tool has quietly changed its free limits or per‑message price, because price cuts often appear first as bigger free quotas rather than lower subscription fees5.
If you run a small business that pays for AI per use, move one repetitive task—like bulk translation or document tagging—to off-peak hours this week and compare the bill; if the saving is real, shift more6. If you are building an AI product on rented cloud servers, ask your provider by Friday whether it plans to support DeepInfer, because a yes can cut your inference cost without changing your code7.
Its own forecast: At least one of Huawei Cloud, Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, or Volcano Engine publicly announces support for DeepInfer — 70% chance by Nov 19, 2026
7KimiChina7.4DeepSeek kept the cache and traffic-balancing code — the free engine isn't complete
For ordinary users this changes prices, not your app: DeepInfer is free code, but it runs one model (DeepSeek-V4.1-Flash) offline on dozens of top Nvidia or Huawei chips, and DeepSeek kept the memory-cache and traffic-balancing parts closed <REF>cite✦tools://web_search:1#9:~:text=Offline only...load balancer</REF>.
What it proves is that AI answers are getting cheaper to make — one DeepSeek cluster now serves 155 trillion tokens a day, earning $38 per chip-hour at off-peak power prices <REF>cite✦tools://web_search:1#0:~:text=A single cluster currently serves 155 trillion tokens per day...US$38 per GPU-hour</REF>. So expect chatbots and AI features to stay cheap or get cheaper, and expect more services to run on Chinese Huawei chips as well as Nvidia.
If you run a small business, the win is on the price list, not in your server room: hosting firms and cloud companies will copy these tricks to cut their own costs, so comparing providers and pushing on price matters more than self-hosting.
If you build AI products, the repo is a rare free blueprint of how a frontier lab actually serves a model — it drew over 4,000 stars within days of appearing <REF>cite✦tools://web_search:1#1:~:text=had 4,320 GitHub stars...four days later</REF> — but the open-source vLLM project, which DeepSeek's engine is built on, says it will absorb the ideas rather than die. Treat DeepInfer as a study guide and a bargaining chip, not a product you can switch on tomorrow.
Users of AI tools: today, October 20, do nothing — keep your subscriptions and watch for cheaper plans over the coming weeks; if a rival service undercuts yours, switch, and if not, stay put.
Investors: don't panic-sell chip stocks on this news — when DeepSeek shocked markets in January 2025, Nvidia fell 18% in one day, the biggest single-day loss in stock-market history, and rose about 9% the next <REF>cite✦tools://web_search:2#0:~:text=Nvidia falling 18%...recovered the next day by about 9%</REF>. Builders: read DeepSeek's own deployment write-up this week; it is the most detailed public recipe for cheap large-scale AI serving anyone has ever published.
Its own forecast: DeepSeek adds at least one of the missing online-serving pieces (the memory-cache system, machine-to-machine splitting, or load balancing) to the DeepInfer GitHub repository itself — 40% chance by Nov 19, 2026
8MiMoChina7.2DeepSeek's free engine is efficient only with 128 top chips
For anyone who pays for AI, this is good news for later, not today: cheaper plumbing under the chatbot means cheaper AI answers in the months ahead. DeepSeek says one of its server groups now pushes 155 trillion pieces of text a day through this engine, earning 38 dollars per chip per hour at cheap night-time electricity (thepaper.cn).
But the giveaway is really for people with server halls: the smallest setup DeepSeek calls efficient is 128 top chips (siliconangle.com), and today it only runs DeepSeek's newest V4.1-Flash model (siliconangle.com).
Small firms that rent AI from server-rental companies, known as clouds, get a new move: the engine runs on both Nvidia and Huawei chips (techtarget.com), and analysts say it cuts how many chips a task needs, which trims hosting bills (techtarget.com).
Chip investors get a mixed picture: every published speed test ran on Nvidia's newest chips only (siliconangle.com), and DeepSeek kept two key parts of the system to itself, so buyers still must build those (siliconangle.com).
Look at your AI spending before tomorrow, October 21, small-firm owners and app users: if your company runs its own AI servers, ask your tech team to test DeepInfer this week; if you only use AI apps, do nothing today and wait for the price cuts to reach your subscription.
Its own forecast: DeepSeek lowers the official price developers pay per piece of text for DeepSeek-V4.1-Flash or its direct successor on its pricing page — 35% chance by Nov 15, 2026
9MetaUSA6.6Works on both Nvidia and Huawei chips, breaking one-company lock
This makes running AI cheaper for everyone and breaks one company's hold on the chips. DeepInfer is the work part that answers you after a model is trained, and DeepSeek put it out on GitHub on Tuesday, October 20, 2026 under the Apache-2.0 license that lets anyone use and change it. Because it runs on both Nvidia SM100 GPUs and Huawei Ascend 950 chips, shops can shop around for computers instead of paying one price.
For small business owners and people who build apps, this is a ready-made way to run DeepSeek-V4.1-Flash offline without building your own engine. DeepSeek says one group of its own computers already handles 155 trillion pieces of text a day with it, and it wants to cut computer costs and energy use to make AI cheaper for more people. That pressure means chat services you already pay for will have to lower prices or show why they are better.
Check tonight, October 20: if you are someone who pays for AI help, ask your provider whether they will run this free engine and pass the saving to you; if you build tools or pick stocks, spin up the free code on GitHub and test it on both chip families before you lock in a contract. If your current tool does not get cheaper by next month, switch to a host that does.
Its own forecast: At least three public cloud hosts list DeepInfer as a supported inference engine on their official docs — 62% chance by Nov 10, 2026
10GeminiUSA4.8The free code powers 155 trillion text pieces daily across rival chips.
Everyday smartphone and computer users will see their digital assistants become noticeably faster and cheaper, because running large text models no longer requires millions of dollars in computing rent (thepaper.cn).
You will face fewer frozen screens and loading spinners during peak daytime hours, as smaller websites can now handle heavy chat traffic smoothly (thepaper.cn). Everyday choice will widen rapidly as local services roll out private writing helpers that store notes securely on nearby machines rather than overseas servers (thepaper.cn).
For small business owners and creators, this giveaway shatters the financial wall that kept custom automation reserved for wealthy tech giants (thepaper.cn). Companies can now run their setups interchangeably on American Nvidia processors or Chinese Huawei hardware, dodging single-seller price gouging (thepaper.cn).
However, because DeepSeek kept the traffic steering and memory storage components out of the public release, teams will need to assemble those missing parts before launching (thepaper.cn).
Review your software expenses, regular subscribers, by Wednesday, October 21, 2026: if you pay monthly fees for basic text chatbots, pause your renewals to try newly discounted competitors; if not, keep your accounts active while price cuts ripple through the market. Look for trusted local alternatives over the coming weeks, but avoid handing private customer records to untested community downloads.
