๐Ÿ”ฎ Will Kimi K3 change the economics of AI?๏ฝœKimi K3 ไผšๆ”นๅ˜ AI ็š„็ปๆตŽๅญฆๅ—๏ผŸ๏ฝœ่‹ฑๆ–‡ๅŽŸๆ–‡ + ไธญๆ–‡็ฟป่ฏ‘

๐Ÿ”ฎ Will Kimi K3 change the economics of AI?๏ฝœKimi K3 ไผšๆ”นๅ˜ AI ็š„็ปๆตŽๅญฆๅ—๏ผŸ๏ฝœ่‹ฑๆ–‡ๅŽŸๆ–‡ + ไธญๆ–‡็ฟป่ฏ‘

ๅฎŒๆ•ดๅ‘ˆ็Žฐ Exponential View 2026 ๅนด 7 ๆœˆ 23 ๆ—ฅๅ…ฌๅผ€ๅฏ่ฏป็š„่‹ฑๆ–‡ๅŽŸๆ–‡ไธŽไธญๆ–‡็ฟป่ฏ‘๏ผŒ่ฟฝ่ธช Kimi K3ใ€token ไปทๆ ผๅผนๆ€งใ€ๅผ€ๆบๆจกๅž‹ๆŽจ็†ๆˆๆœฌๅ’Œไธญๅ›ฝ AI ๅฎž้ชŒๅฎคๆ•ˆ็އ๏ผ›ไป˜่ดนๅข™ๅŽ็š„ๅ†…ๅฎนไธ่กฅๅ†™ใ€‚

English original

๐Ÿ”ฎ Will Kimi K3 change the economics of AI?

The pressure is on

Jul 23, 2026
โˆ™ Paid
Moonshot AI office interior
From our visit to Moonshot AIโ€™s office earlier this year 1
Kimi K3 has caused quite an uproar since its release last week. Itโ€™s the first time a Chinese model has taken the lead on the frontend Code Arena benchmark. And thatโ€™s three months since Moonshot AIโ€™s previous impressive flagship model, Kimi K2.6, was released. 1
Following in Moonshotโ€™s steps, Alibaba announced over the weekend that Qwen3.8 โ€“ a 2.4 trillion-parameter model โ€“ is coming soon, and unlike its last release, this one will be an open-weight model. No benchmarks or further details have been released as of yet. 1
Open models are now estimated to be 4-7 months behind the frontier in cyber capabilities, down from 6-10 months in 2025. And despite compute constraints, efficiency improvements mean these labs are doing more with less. Comparing the compute availability and model performance between US labs and Chinese labs, we estimated Chinese labs to be getting 4-7x more out of their compute. 1
Hannah Petrovic and I spent some time with the Moonshot AI and Alibaba teams in China back in April and May, and weโ€™ve had time to think about the economics of open-source models and how they affect the entire ecosystem. 1

Does Kimi K3 break the economic case for AI?

Some have claimed that Kimi K3โ€™s performance breaks the economic case for AI as it lowers the cost to complete various tasks at frontier standards. For instance, Microsoft engineers are reportedly testing whether Kimi K3 can be used within Copilot. 1
We donโ€™t think this is the case, and in todayโ€™s post weโ€™ll work through what might happen next. 1
In The State of the AI Economy report, we found that token usage is elastic across providers. This means that every drop in token price leads to a larger increase in token volume, more than offsetting the difference. 1
Token price and consumption elasticity chart from the original article
Price declines track higher token volumes in the sourceโ€™s comparison. 1
For every 10% price cut, token consumption rises 12-18%. A paper by Demirer et al, found a similar effect: a 10% price cut resulted in an 11% or so increase in volumes, which economists call an elasticity of -1.11. 1
The net effect is a rise in total token spend. But note that the effect is a weak one, not the cantering Jevonsโ€™ paradox sometimes presented. Reality might tilt the scales further in favor of more, not less, demand. Workflows are becoming more token-intensive as we rely on reasoning models and verification and approval loops. And the early evidence suggests that firms that adopt AI early tend to increase their relative spend alongside growing headcount. These effects might be short-term elasticities rather than ones that can be sustained for decades, but for now they indicate that falling prices increase volumes and, with that, revenue. 1

Flowing down the stack

The model weights may be free, but the inference is not. Kimi K3 has 2.8 trillion parameters. The weights alone occupy 1.4 TB. It needs to be served on something like a 72-GPU NVIDIA GB200 NVL72 rack or equivalent. Thatโ€™ll cost $3-4 million to buy and install. Operating it consumes about 120 kW continuously, over a million kWh per year, before you consider networking, storage, cooling, and humans. If you rented these in the open market, it would cost about $7 million a year. 1
But for infrastructure providers, the economics of hosting open-source models can be very attractive compared to serving closed-source models. A simple way to understand this is to think of the hyperscaler as needing to pay a license fee for a closed-source model but not for an open-source one1. 1
Public access boundary
The publicly readable page ends here with the notice "This post is for paid subscribers." The remaining article is not accessible on the official page and is not translated or supplemented.

ไธญๆ–‡็ฟป่ฏ‘

๐Ÿ”ฎ Kimi K3 ไผšๆ”นๅ˜ AI ็š„็ปๆตŽๅญฆๅ—๏ผŸ

ๅŽ‹ๅŠ›ๆญฃๅœจๅŠ ๅคง

2026 ๅนด 7 ๆœˆ 23 ๆ—ฅ
โˆ™ ไป˜่ดนๆ–‡็ซ 
ไปŠๅนดๆ—ฉไบ›ๆ—ถๅ€™ๆˆ‘ไปฌ่ฎฟ้—ฎ Moonshot AI ๅŠžๅ…ฌๅฎคๆ—ถๆ‹ๆ‘„็š„็…ง็‰‡ 1
Kimi K3 ไธŠๅ‘จๅ‘ๅธƒๅŽๅผ•่ตทไบ†ไธๅฐ็š„ๅๅ“ใ€‚่ฟ™ๆ˜ฏไธญๅ›ฝๆจกๅž‹้ฆ–ๆฌก็™ปไธŠ Frontend Code Arena ๅŸบๅ‡†ๆต‹่ฏ• ็š„ๆฆœ้ฆ–ใ€‚Moonshot AI ไธŠไธ€ๆฌพ่กจ็ŽฐๅŒๆ ทไบฎ็œผ็š„ๆ——่ˆฐๆจกๅž‹ Kimi K2.6๏ผŒๅˆ™ๆ˜ฏๅœจไธ‰ไธชๆœˆๅ‰ๅ‘ๅธƒ็š„ใ€‚1
้˜ฟ้‡Œๅทดๅทดๆฒฟ็€ Moonshot ็š„ๆญฅไผ๏ผŒๅœจๅ‘จๆœซๅฎฃๅธƒ๏ผŒQwen3.8๏ผŒไธ€ไธชๆ‹ฅๆœ‰ 2.4 ไธ‡ไบฟๅ‚ๆ•ฐ็š„ๆจกๅž‹๏ผŒๅณๅฐ†ๆŽจๅ‡บใ€‚ไธŽไธŠไธ€ๆฌกๅ‘ๅธƒไธๅŒ๏ผŒ่ฟ™ๆฌกไผšๆ˜ฏไธ€ไธชๅผ€ๆ”พๆƒ้‡ๆจกๅž‹ใ€‚ๆˆช่‡ณ็›ฎๅ‰๏ผŒๅฎ˜ๆ–น่ฟ˜ๆฒกๆœ‰ๅ‘ๅธƒๅŸบๅ‡†ๆˆ็ปฉๆˆ–ๆ›ดๅคš็ป†่Š‚ใ€‚1
็›ฎๅ‰ไผฐ่ฎก๏ผŒๅผ€ๆ”พๆจกๅž‹ๅœจ็ฝ‘็ปœๅฎ‰ๅ…จ่ƒฝๅŠ›ไธŠ่ฝๅŽๅ‰ๆฒฟๆจกๅž‹ 4 ่‡ณ 7 ไธชๆœˆ๏ผŒ2025 ๅนด่ฟ™ไธ€ๅทฎ่ทๆ˜ฏ 6 ่‡ณ 10 ไธชๆœˆใ€‚ๅฐฝ็ฎก็ฎ—ๅŠ›ๅ—้™๏ผŒๆ•ˆ็އๆๅ‡่ฎฉ่ฟ™ไบ›ๅฎž้ชŒๅฎค็”จๆ›ดๅฐ‘่ต„ๆบๅšๆ›ดๅคšไบ‹ๆƒ…ใ€‚ๆฏ”่พƒ็พŽๅ›ฝไธŽไธญๅ›ฝๅฎž้ชŒๅฎค็š„ๅฏ็”จ็ฎ—ๅŠ›ๅ’Œๆจกๅž‹่กจ็ŽฐๅŽ๏ผŒๆˆ‘ไปฌไผฐ่ฎกไธญๅ›ฝๅฎž้ชŒๅฎคไปŽ็ฎ—ๅŠ›ไธญๅพ—ๅˆฐ็š„ไบงๅ‡บๅคš 4 ่‡ณ 7 ๅ€ใ€‚1
ไปŠๅนด 4 ๆœˆๅ’Œ 5 ๆœˆ๏ผŒๆˆ‘ไธŽ Hannah Petrovic ๅœจไธญๅ›ฝๅ’Œ Moonshot AIใ€้˜ฟ้‡Œๅทดๅทด็š„ๅ›ข้˜Ÿไบคๆตไบ†ไธ€ๆฎตๆ—ถ้—ด๏ผŒไนŸๆœ‰ๆ—ถ้—ดๆ€่€ƒๅผ€ๆบๆจกๅž‹็š„็ปๆตŽๅญฆ๏ผŒไปฅๅŠๅฎƒไปฌๅฆ‚ไฝ•ๅฝฑๅ“ๆ•ดไธช็”Ÿๆ€็ณป็ปŸใ€‚1

Kimi K3 ไผšๆ‰“็ ด AI ็š„็ปๆตŽๅŸบ็ก€ๅ—๏ผŸ

ๆœ‰ไบบ่ฎคไธบ๏ผŒKimi K3 ๆŒ‰ๅ‰ๆฒฟๆ ‡ๅ‡†ๅฎŒๆˆๅ„็งไปปๅŠก็š„ๆˆๆœฌไธ‹้™๏ผŒๅ› ๆญคๅฎƒ็š„่กจ็Žฐๆ‰“็ ดไบ† AI ็š„็ปๆตŽๅŸบ็ก€ใ€‚็›ธๅ…ณ่ฎจ่ฎบๅŒ…ๆ‹ฌๆˆๆœฌไธ‹้™ไปฅๅŠไปฅๅ‰ๆฒฟๆฐดๅนณๅฎŒๆˆไปปๅŠก๏ผ›ๆฎๆŠฅ้“๏ผŒๅพฎ่ฝฏๅทฅ็จ‹ๅธˆๆญฃๅœจๆต‹่ฏ• Kimi K3 ๆ˜ฏๅฆ่ƒฝ็”จไบŽ Copilotใ€‚1
ๆˆ‘ไปฌไธ่ฟ™ไนˆ่ฎคไธบใ€‚ๆœฌๆ–‡ๅฐ†ๆขณ็†ๆŽฅไธ‹ๆฅๅฏ่ƒฝๅ‘็”Ÿไป€ไนˆใ€‚1
ๅœจใ€ŠAI ็ปๆตŽ็Žฐ็Šถใ€‹ๆŠฅๅ‘Šไธญ๏ผŒๆˆ‘ไปฌๅ‘็Žฐ๏ผŒไธๅŒๆœๅŠกๅ•†ไน‹้—ด็š„ token ไฝฟ็”จ้‡ๅ…ทๆœ‰ๅผนๆ€งใ€‚่ฟ™ๆ„ๅ‘ณ็€ token ไปทๆ ผๆฏไธ‹้™ไธ€็‚น๏ผŒtoken ไฝฟ็”จ้‡ๅฐฑไผšๆ›ดๅคงๅน…ๅขžๅŠ ๏ผŒๅขž้‡่ถณไปฅๆŠตๆถˆไปทๆ ผๅทฎๅผ‚ใ€‚1
ๆฏๆฌก้™ไปท 10%๏ผŒtoken ๆถˆ่€—้‡ไผšไธŠๅ‡ 12% ่‡ณ 18%ใ€‚Demirer ็ญ‰ไบบ็š„่ฎบๆ–‡ๅ‘็Žฐไบ†็ฑปไผผๆ•ˆๅบ”๏ผšไปทๆ ผไธ‹้™ 10% ไผšๅธฆๆฅ็บฆ 11% ็š„ๆ•ฐ้‡ๅขž้•ฟ๏ผŒ็ปๆตŽๅญฆๅฎถๆŠŠ่ฟ™็งฐไธบ -1.11 ็š„ๅผนๆ€งใ€‚1
ๅ‡€ๆ•ˆๆžœๆ˜ฏ token ๆ€ปๆ”ฏๅ‡บไธŠๅ‡ใ€‚ไฝ†่ฆๆณจๆ„๏ผŒ่ฟ™ไธชๆ•ˆๅบ”ๅนถไธๅผบ๏ผŒไธๆ˜ฏๆœ‰ๆ—ถ่ขซๆ่ฟฐ็š„้‚ฃ็งๅฅ”่…พๅผ Jevons ๆ‚–่ฎบใ€‚็Žฐๅฎžๅฏ่ƒฝ่ฟ˜ไผš่ฟ›ไธ€ๆญฅๅๅ‘ๆ›ดๅคšใ€่€Œไธๆ˜ฏๆ›ดๅฐ‘็š„้œ€ๆฑ‚ใ€‚้š็€ๆˆ‘ไปฌไพ่ต–ๆŽจ็†ๆจกๅž‹ใ€้ชŒ่ฏไธŽๅฎกๆ‰นๅพช็Žฏ๏ผŒๅทฅไฝœๆตๆญฃๅœจๆถˆ่€—ๆ›ดๅคš tokenใ€‚ๆ—ฉๆœŸ่ฏๆฎ่กจๆ˜Ž๏ผŒ่พƒๆ—ฉ้‡‡็”จ AI ็š„ไผไธšๅœจๅ‘˜ๅทฅไบบๆ•ฐๅขž้•ฟ็š„ๅŒๆ—ถ๏ผŒๅพ€ๅพ€ไนŸไผšๅขžๅŠ ็›ธๅฏนๆ”ฏๅ‡บใ€‚่ฟ™ไบ›ๆ•ˆๅบ”ๅฏ่ƒฝๅชๆ˜ฏ็ŸญๆœŸๅผนๆ€ง๏ผŒๆ— ๆณ•็ปดๆŒๆ•ฐๅๅนด๏ผŒไฝ†็›ฎๅ‰ๅฎƒไปฌ่กจๆ˜Ž๏ผŒไปทๆ ผไธ‹้™ไผšๅขžๅŠ ไฝฟ็”จ้‡๏ผŒๅนถ้šไน‹ๅขžๅŠ ๆ”ถๅ…ฅใ€‚1

ๆˆๆœฌๆฒฟ็€ๆŠ€ๆœฏๆ ˆๅ‘ไธ‹ไผ ๅฏผ

ๆจกๅž‹ๆƒ้‡ๅฏไปฅๅ…่ดน๏ผŒๆŽจ็†ๅดไธๆ˜ฏใ€‚Kimi K3 ๆœ‰ 2.8 ไธ‡ไบฟไธชๅ‚ๆ•ฐ๏ผŒไป…ๆƒ้‡ๅฐฑๅ ๆฎ 1.4 TB ็š„็ฉบ้—ดใ€‚ๅฎƒ้œ€่ฆ้ƒจ็ฝฒๅœจ็ฑปไผผไธ€ๅฐ้…ๅค‡ 72 ๅ— GPU ็š„ NVIDIA GB200 NVL72 ๆœบๆžถไธŠ๏ผŒๆˆ–ไฝฟ็”จๅŒ็ญ‰่ฎพๅค‡ใ€‚่ดญไนฐๅ’Œๅฎ‰่ฃ…่ฟ™ๆ ท็š„่ฎพๅค‡่ฆ่Šฑ 300 ไธ‡่‡ณ 400 ไธ‡็พŽๅ…ƒใ€‚่ฎพๅค‡ๆŒ็ปญ่ฟ่กŒๆ—ถ็š„ๅŠŸ่€—็บฆไธบ 120 kW๏ผŒๆฏๅนด่ถ…่ฟ‡ 100 ไธ‡ kWh๏ผ›่ฟ™่ฟ˜ๆฒกๆœ‰็ฎ—ๅ…ฅ็ฝ‘็ปœใ€ๅญ˜ๅ‚จใ€ๅ†ทๅดๅ’Œไบบๅ‘˜ๆˆๆœฌใ€‚ๅฆ‚ๆžœๅœจๅ…ฌๅผ€ๅธ‚ๅœบ็งŸ็”จ่ฟ™ๆ ท็š„่ฎพๅค‡๏ผŒๆฏๅนดๆˆๆœฌ็บฆไธบ 700 ไธ‡็พŽๅ…ƒใ€‚1
ไฝ†ๅฏนๅŸบ็ก€่ฎพๆ–ฝๆไพ›ๅ•†ๆฅ่ฏด๏ผŒไธŽๆไพ›้—ญๆบๆจกๅž‹ๆœๅŠก็›ธๆฏ”๏ผŒๆ‰˜็ฎกๅผ€ๆบๆจกๅž‹็š„็ปๆตŽๆ€งๅฏ่ƒฝๅพˆๆœ‰ๅธๅผ•ๅŠ›ใ€‚ๅฏไปฅ่ฟ™ๆ ท็†่งฃ๏ผš่ถ…ๅคง่ง„ๆจกไบ‘ๆœๅŠกๅ•†้œ€่ฆไธบ้—ญๆบๆจกๅž‹ๆ”ฏไป˜่ฎธๅฏ่ดน๏ผŒๅดไธ็”จไธบๅผ€ๆบๆจกๅž‹ๆ”ฏไป˜่ฎธๅฏ่ดน1ใ€‚1
ๅ…ฌๅผ€ๅ†…ๅฎน่พน็•Œ
ๅฎ˜ๆ–น้กต้ขๅœจไธŠไธ€ไธชๆฎต่ฝๅŽๆ˜พ็คบใ€ŒThis post is for paid subscribersใ€ใ€‚ๅ…ถไฝ™ๆ–‡็ซ ๅ†…ๅฎนๅœจๅฝ“ๅ‰้กต้ขไธŠไธๅฏ่ฎฟ้—ฎ๏ผŒๆœฌๆ–‡ไธ็ฟป่ฏ‘๏ผŒไนŸไธ่กฅๅ†™ใ€‚

๊ด€๋ จ ์ฝ˜ํ…์ธ 

  • ๋กœ๊ทธ์ธํ•˜๋ฉด ๋Œ“๊ธ€์„ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
More from this channel