이 페이지는 영어 원문을 번역한 것이며, 내용에 차이가 있을 경우 영어 원문이 우선합니다. English
주요 발견
2026년 7월 Steam 하드웨어 설문에서 조사 대상 시스템의 25.9%가 그래픽 메모리 16 GB, 5.4%가 24 GB를 갖추고 있었습니다. [S1]
필자의 테스트에서 24 GB Mac은 모델을 불러오기도 전인 시작 직후에 시스템만으로 9.67 GB의 메모리를 사용했습니다. [S22]
필자가 테스트한 로컬 AI 앱 아홉 개 중 여덟 개는 점수를 매길 수 있었고, 그 여덟 개 중 네 개는 핵심 기능에서 2점 이하를 받았습니다. [S22]
카드 결제 데이터를 다룬 세 출처에 따르면, 미국 가구 또는 소비자의 2.2%에서 4.5%가 AI 챗봇에 비용을 냅니다. 상위 1% 결제자가 소비자 AI 지출의 19.5%를 차지합니다. [S11, S12, S13]
128 GB NVIDIA DGX Spark의 정가는 출시 당시 $3,999에서 2026년 2월 $4,699, 2026년 10월 $6,950으로 올랐습니다. [S3, S4]
2026년 3분기에 신규 노트북과 데스크톱 판매가 감소했습니다. IDC는 출하량이 20.1% 줄어 6,270만 대, Omdia는 21.2% 줄어 5,810만 대라고 보고했습니다. [S43]
가장 큰 오픈 모델은 최고의 클라우드 모델에 약 4개월 뒤처져 있지만, 최소 377 GB의 메모리가 필요합니다(계산값). [S14, S19]
"local AI"를 검색하면 Apple의 독일 App Store에서 10건, Google Play에서 29건이 나옵니다. 전체 앱 수는 각각 약 275만 개와 265만 개입니다. [S34, S35, S36]
Google에서 "local ai" 검색은 2026년 5월 말에 정점을 찍었고 10월 중순에는 1년 전의 절반 수준이었습니다. 보고서는 이것이 다른 지표와 엇갈리는 이유를 데이터만으로는 설명할 수 없다고 밝힙니다. [S29]
프레스 키트: 핵심 수치와 인용문기자를 위한 자료입니다. 정확한 출처 표기가 붙은 수치 일곱 개, 바로 복사해 쓸 수 있는 인용문, 저자와 방법 설명을 담았습니다.
핵심 수치
5.4% 2026년 7월 Steam 하드웨어 설문에서 시스템의 5.4%가 그래픽 메모리 24 GB를 갖추고 있었고, 25.9%는 16 GB였습니다. 출처: Steam Hardware & Software Survey, 2026년 7월, 2차 보도를 통해 확인(Valve 1차 페이지는 확인하지 못함) [S1]
8개 중 4개 필자가 점수를 매길 수 있었던 로컬 AI 앱 중 핵심 기능에서 2점 이하를 받은 앱입니다(아홉 개를 테스트했고 여덟 개가 점수를 받았습니다). 출처: 필자의 직접 테스트, 2026년 10월 9일~24일 [S22]
9.67 GB 모델을 불러오기 전, 시작 직후 24 GB Mac에서 사용 중이던 메모리입니다. 출처: 필자의 직접 테스트, 24 GB Mac [S22]
$3,999에서 $6,950 128 GB NVIDIA DGX Spark의 정가로, 출시 당시부터 2026년 10월까지의 변화입니다(2026년 2월에는 $4,699). 출처: NVIDIA 개발자 포럼(1차), Implicator.ai(2차) [S3, S4]
2.2%에서 4.5% 미국 가구 또는 소비자 중 AI 챗봇에 비용을 내는 비율입니다. 상위 1% 결제자가 소비자 AI 지출의 19.5%를 차지합니다. 출처: Bank of America Institute, PNC(TechDogs 경유), a16z/YipitData(AI Weekly 경유) [S11, S12, S13]
-20.1% / -21.2% 2026년 3분기 PC 출하량 증감률로, IDC는 6,270만 대, Omdia는 5,810만 대라고 보고했습니다. 출처: IDC와 Omdia, Digital Today와 Windows Latest 보도 인용(2차) [S43]
10건과 29건 Apple의 독일 App Store와 Google Play에서 "local AI"를 검색한 결과 수이며, 전체 앱 수는 각각 약 275만 개와 265만 개입니다. 출처: 2026년 10월 23일에 캡처한 스토어 검색 결과, 42matters 스토어 통계 [S34, S35, S36]
복사용 인용문
오늘날 로컬 AI에는 닭이 몇 마리 안 되고 달걀은 더 적습니다. 2026년 7월 Steam 설문에서 그래픽 메모리 24 GB를 갖춘 시스템은 5.4%에 불과했고, 제가 점수를 매길 수 있었던 로컬 AI 앱 여덟 개 중 네 개는 핵심 기능에서 2점 이하를 받았습니다. (Hans Kuepper, 로컬 AI 현황, 보고서 01, Q4 2026)
NVIDIA의 128 GB DGX Spark 정가는 출시 당시 $3,999에서 2026년 10월 $6,950으로 올랐습니다. 로컬 AI를 편하게 쓰게 해 주는 하드웨어는 싸지는 것이 아니라 비싸지고 있습니다. (Hans Kuepper, 로컬 AI 현황, 보고서 01, Q4 2026)
저자와 방법에 대하여
Hans Kuepper는 프랑크푸르트에 기반을 둔 PromptQuorum의 창업자로, 30년 넘게 컨설팅과 경영 일선에서 일해 왔습니다. 이 보고서의 앱 아홉 개는 직접 테스트했습니다.
수치는 번호가 붙은 출처에서 가져왔고, 각 주장에는 증거 표시가 붙으며, 가정에는 누가 그렇게 보는지와 무엇에 근거하는지를 밝혔습니다. 2차 출처는 그렇게 표시했습니다. 주장 목록과 모든 출처는 내려받을 수 있습니다.
쉬운 설명판 · 작성자 Hans Kuepper · 2026년 10월 23일 · 로컬 AI 생태계인 PromptQuorum를 개발해 온 경험에 바탕을 둔 필자 개인의 시각
로컬 AI는 아직 갈 길이 멉니다. 좋은 모델을 돌릴 수 있는 컴퓨터는 극히 적고, 애써 써볼 만한 앱은 그보다 더 적습니다. 저렴하고, 프라이버시를 지키고, 대형 클라우드 모델만큼 뛰어나다는 약속도 아직은 일반 사용자가 갈아탈 이유가 되지 못합니다. 필자는 이를 “닭은 몇 마리 안 되고, 달걀은 그보다 더 적다”라고 표현합니다.
닭: 턱없이 부족한 하드웨어
전체 노트북과 데스크톱 PC 가운데 좋은 로컬 모델을 돌릴 수 있는 비율은 어디에도 공개되어 있지 않습니다. 가장 가까운 대용 지표는 Valve의 Steam 설문입니다. 대상이 일반 사용자보다 장비가 좋은 게이머라는 점은 감안해야 합니다. 로컬 AI 모델은 주로 그래픽카드의 메모리에서 실행되며, 이는 컴퓨터의 메인 메모리와 별개입니다. PC에 메인 메모리가 32 GB 있어도 그래픽카드에는 8 GB나 16 GB뿐일 수 있고, 모델은 그래픽카드 메모리에 들어가야 합니다. 게이머 사이에서도 2026년 7월 기준 그래픽 메모리 16 GB를 갖춘 시스템은 25.9%, 24 GB는 5.4%에 불과했습니다.[S1] Apple의 Mac은 방식이 다릅니다. 하나의 공유 메모리 풀이 두 역할을 모두 맡기 때문에 24 GB Mac으로도 테스트가 가능하며, Mac이 시스템의 1.9%를 차지하는 Steam 설문이 Mac에 대해 알려 주는 바가 적은 것도 이 때문입니다.[S2] 사양상의 메모리가 곧 쓸 수 있는 메모리인 것도 아닙니다. 필자의 테스트에서 24 GB Mac은 시작 직후 시스템만으로 9.67 GB를 차지했습니다.[S22]
NVIDIA DGX Spark (128 GB), 정가(USD)
출처: 보도된 NVIDIA 가격 발표(S3, S4). 10월 수치는 근거 자료가 상대적으로 부족합니다. 출시 대비 상승률: +74% (계산값).
대형 모델을 돌릴 수 있는 하드웨어도 메모리 부족 때문에 저렴해지기는커녕 오히려 비싸지고 있습니다.[S6, S7] 2026년 3분기에는 신규 노트북과 데스크톱 판매가 감소했습니다. IDC는 출하량이 20.1% 줄어 6,270만 대, Omdia는 21.2% 줄어 5,810만 대라고 보고했으며, 둘 다 메모리 부족을 원인으로 지목합니다.[S43] 따라서 성능을 갖춘 기기의 기반은 작고, 진입 가격은 오르고 있습니다.
신규 기기 중 “AI 지원” 기기의 비율은 얼마이며, 그것은 무슨 의미일까요? Gartner는 2025년 출하 노트북과 PC의 31%, 2026년에는 54.7%가 “AI PC”가 될 것으로 예측했습니다. 이 예측은 부족 사태 이전인 2025년 8월 자료입니다(1년 이상 지난 자료).[S8] 2026년의 실측 비율은 찾지 못했습니다. 확인된 가장 최근의 실측 비율은 Canalys의 2024년 2분기 14%, 3분기 20%입니다(1년 이상 지난 자료).[S44] “AI 지원”은 칩 기준의 정의입니다. Canalys는 신경망 처리 장치(NPU) 같은 AI 전용 칩셋을 탑재한 노트북이나 데스크톱을 집계하며, Microsoft의 Copilot+ PC는 초당 40조 회 이상 연산하는 NPU, 16 GB 메모리, 256 GB 저장 공간을 요구합니다.[S44, S45] 확인한 정의 중 대규모 언어 모델의 실행을 기준으로 삼은 것은 하나도 없습니다.
달걀: 써볼 만한 앱은 더욱 적음
필자는 로컬 AI 앱 아홉 개를 테스트했습니다. 여덟 개는 24 GB Mac에서, 한 개는 Android 휴대전화에서 테스트했습니다. 점수를 매길 수 있었던 여덟 개 중 네 개는 핵심 기능을 제대로 수행하지 못했습니다.[S22] 실패의 원인은 인터페이스인 경우가 많습니다. 앱들은 모델의 크기도, 내 컴퓨터에서 실행 가능한지도 알려 주지 않은 채 모델 목록만 보여 줍니다. 한 앱은 Mac의 메모리보다 큰 다운로드를 권했고, 다른 앱은 총 129 GB의 다운로드 목록을 보여 주었는데 주 모델 하나(41.5 GB)만으로도 Mac의 메모리 24 GB보다 크지만 눈에 띄는 메모리 경고는 없었습니다. 네 개 앱은 다운로드 남은 시간을 표시하지 않았습니다. 한 앱은 어떤 모델을 설치했는지 끝내 알려 주지 않았습니다.[S22] 테스트한 앱 가운데 가장 완성도가 높았던 LU는 각 모델의 크기와 Mac에서의 실행 가능 여부를 보여 주었고, 코딩 에이전트는 실제로 동작하는 게임을 만들어 냈습니다. 다만 문서 채팅은 PDF, Excel, Word 파일을 읽지 못했습니다.[S22]
쓸 만한 것은 아직 거의 없습니다. Apple App Store에는 약 275만 개, Google Play에는 약 265만 개의 앱이 있습니다.[S36] “local AI”를 검색하면 Apple의 독일 스토어에서는 10건, Google Play에서는 29건이 나옵니다.[S34, S35] 수백만 개 중의 숫자입니다. 평점이 표시된 경우 2.2점에서 4.2점 사이입니다.
존재하는 것들 상당수는 일반 사용자가 아니라 개발자를 위해 만들어진 플랫폼인 GitHub에 있습니다. 필자의 테스트 기록 아홉 건 중 최소 다섯 건(Page Assist, LTX Desktop, Unsloth, DiffusionBee, BoBe)에서 GitHub가 공급사 페이지, README 또는 릴리스 출처였습니다. Page Assist의 README는 터미널과 Bun이라는 도구가 필요한 소스 빌드를 맨 앞에 안내하고 있으며, 테스터의 첫 시도는 실패했습니다.[S22] 게다가 양도 많습니다. GitHub에는 “ollama” 토픽의 공개 저장소가 5,794개, “local-llm” 토픽에는 9,477개가 있습니다.[S46] 상당수는 서로의 포크이거나 복사본입니다. 필자의 표현으로는, 무엇이 최신이고 무엇이 관리되고 있으며 무엇이 안전한지 아무도 알 수 없는 혼란 상태입니다. 일반 사용자에게 GitHub에만 있는 앱은 사실상 구할 수 없는 앱입니다.
약속: 저렴하고, 프라이버시를 지키고, 최첨단 수준일 것
저렴하다. 비용 절감은 AI에 돈을 쓰는 사람에게만 의미가 있습니다. 가장 앞선 시장인 미국에서도 AI 챗봇에 비용을 내는 가구 또는 소비자는 2.2%에서 4.5% 사이입니다.[S11, S12, S13] 나머지 사람들에게 클라우드 도구는 이미 무료이므로 로컬 앱이 가격으로 앞설 여지가 없습니다. 구독료를 아끼려면 구매자가 먼저 위에서 본 하드웨어 비용을 치러야 합니다.
프라이버시를 지킨다. 프라이버시는 가장 강력한 약속이지만 판매 논거로는 가장 약합니다. 한 설문에서 챗봇 이용자의 32%는 친구, 가족, 의사에게도 털어놓지 않은 이야기를 챗봇에 했고, 53%는 대화가 AI 학습에 사용될 수 있다는 사실을 몰랐거나 확실히 알지 못했습니다.[S20] 20년 동안 사람들은 Google 등이 제공하는 우수한 무료 서비스를 위해 프라이버시를 내주었고, 그 서비스보다 못한 로컬 앱이 이를 바꾸지는 못할 것입니다. 앱 자체가 좋아지기 전까지 로컬 AI의 프라이버시 약속은 대체로 마케팅에 가깝습니다.
최첨단 수준이다. 흔히 내세우는 주장은 오픈 모델이 최고의 클라우드 모델에 몇 달밖에 뒤지지 않는다는 것입니다. 가장 큰 오픈 모델은 실제로 약 4개월 뒤처져 있지만,[S14] 최소 377 GB의 메모리가 필요하며(계산값),[S19] 이는 어떤 소비자용 기기로도 감당할 수 없는 수준입니다. 게이밍 그래픽카드 한 장에 들어가는 모델은 Epoch AI에 따르면 6개월에서 12개월 뒤처지며, Epoch AI는 실제 격차는 아마 더 클 것이라고 덧붙입니다(1년 이상 지난 자료).[S15] 노트북이나 휴대전화에 대한 수치는 없습니다. 필자의 테스트에서 Android 휴대전화의 1 GB 모델은 현직 국가원수가 암살되었다고 답했는데, 제3의 출처는 이를 반박합니다.[S22] “몇 달 차이”라는 수치는 마케팅 논거입니다. 사용자는 이런 앱에서 최첨단 수준의 결과를 기대하지만 얻지 못합니다. 인터페이스는 복잡하고, 설정은 복잡하며, 결과는 부실하거나 아예 나오지 않습니다. 필자의 테스트에서는 한 앱이 설정 단계에서 멈춰 아무것도 만들어 내지 못했고, 한 앱은 첫 그림을 만들기 전에 메모리가 부족해졌으며, 한 앱은 잘못된 모델을 선택한 뒤 회색 사각형을 출력했습니다.[S22]
자금: 투자는 극히 미미함
자본은 시장을 따라 움직이는데, 아직 들어오지 않았습니다. 모델을 실행하는 소프트웨어인 Ollama는 2026년 7월에 6,500만 달러를 조달했습니다.[S17] 그 위에서 동작하는 앱의 경우, 웹 검색으로 찾은 것은 GPT4All 개발사의 오래된 투자 라운드(2023년)와 AnythingLLM 개발사에 관한 작은 기록 하나뿐이며, 이름을 확인한 다른 앱 다섯 개에 대해서는 아무것도 찾지 못했습니다.[S27, S28] 벤처 투자 라운드는 보통 공개되므로 필자는 이를 투자가 극히 적다는 뜻으로 해석합니다. 필자가 연락하고 있는 한 앱은 최근 투자자를 찾았습니다(실명은 밝히지 않음).
아직은 갈 길이 멉니다
Google Trends, 전 세계, 주간: "local ai"와 "local llm" (지수 0-100)
출처: 필자가 제공한 Google Trends CSV(S29). 지수는 상대값이며, 100은 해당 기간 두 검색어 중 가장 높은 주간 값입니다. 12개월치만 제공되므로 계절성은 확인할 수 없습니다.
Google에서 “local ai” 검색은 2026년 5월 말에 정점을 찍은 뒤, 10월 중순에는 1년 전의 절반 수준으로 떨어졌습니다.[S29] 이 하락은 7월 어느 한 주에 계단식으로 일어났습니다. 무슨 일이 벌어지고 있는지는 데이터만으로는 설명할 수 없습니다. 다른 지표는 상승을 가리키는데 Google Trends만 하락을 가리킵니다. Ollama는 1월부터 7월 사이에 개발자 기반이 두 배로 늘었다고 밝혔고,[S17] 필자는 다른 트렌드는 계속 상승 중이라고 전합니다(8월부터 10월까지는 여기서 검증하지 않았습니다).
로컬 AI는 개발자 사이에서는 성장하고 있습니다. 하지만 일반 사용자가 갈아탈 이유는 아직 없습니다. 하드웨어는 너무 적고, 좋은 앱은 너무 적으며, 약속은 실제 사용에서 지켜지지 않고, 이를 바꿀 자본도 부족하기 때문입니다.
요약: 이제 막 시작입니다
로컬 AI는 이제 막 시작 단계입니다. 개발자 사이에서는 잘 알려져 있습니다. Ollama의 개발자 기반은 6개월 만에 두 배가 되었고, 중국 연구소들은 2026년 내내 거의 매달 최대 규모의 오픈 모델을 내놓았습니다.[S17, S19] 일반 대중에게는 거의 알려져 있지 않습니다. 대부분의 사람은 무엇이 가능한지조차 모릅니다.
그런데 막상 써 보면 어떨까요? 먼저 하드웨어의 벽에 부딪힙니다. 노트북이나 PC의 사양이 충분하지 않고, 게이밍 PC 중에서도 좋은 모델에 필요한 그래픽 메모리를 갖춘 비율은 얼마 되지 않습니다.[S1] 그다음 앱을 쓰기 시작하면, 클라우드에서 접해 온 최첨단 모델에 비해 대부분 기대에 못 미칩니다.[S22]
그래서 이 시장은 모든 신기술 시장이 그렇듯 얼리 어답터와 애호가가 주도하고 있습니다. 이 상황은 빠르게 바뀌고 있습니다. 모델은 크기 대비 성능이 계속 좋아지고 있습니다.[S37]
국가별로도 같은 패턴이 나타나는 것이 보통입니다. 신기술은 선진 경제권에서 먼저 시작되고 나중에 나머지 지역으로 퍼지는 경향이 있습니다.[S31] 필자의 인상으로는 미국과 중국이 로컬 AI에서 가장 앞서 있습니다. 대부분의 AI 모델이 그곳에서 만들어지고 가장 큰 오픈 모델은 중국 연구소에서 나오기 때문입니다.[S19] 반면 다른 대부분의 국가는 이제 막 시작한 단계입니다. 로컬 AI에 대한 국가별 수치는 존재하지 않으며, AI 이용 전반에 관한 유일한 국가별 데이터는 국가 간 큰 격차를 보여 줍니다.[S30] 필자는 이 흐름이 전 세계로 확산될 것으로 예상합니다.
하드웨어도 뒤따르겠지만, 대가가 따릅니다. 컴퓨터는 갈수록 AI 지원 기능을 갖추겠지만, 얼마나 쓸 만한지는 가격에 달려 있습니다. 이것이 핵심입니다. 메모리 64 GB 이상의 기기가 대량으로 팔리려면 가격이 크게 내려가야 하는데, 전망은 그렇게 될지에 대해 엇갈립니다. Counterpoint는 새 메모리 생산 능력이 2028년까지 가격을 낮출 것으로 보지만,[S38] TrendForce는 소비자용 메모리 가격이 내년에도 계속 오를 것으로 보고,[S38] IDC는 부족 사태가 2027년까지 이어질 것으로 봅니다.[S7]
앱도 뒤따를 것입니다. Appfigures에 따르면 신규 앱 출시는 2026년 1분기에 전년 동기 대비 60%, 4월 현재까지는 104% 늘었습니다.[S41] 이 급증의 배경에는 앱 스토어와 GitHub 모두에서 AI 코딩 도구가 있습니다. 창의성은 개인에게로 옮겨 가고 있으며, 개인은 몇 년 전에는 할 수 없던 일을 오늘 할 수 있습니다. 개인은 로컬 AI로 작업하며 자신의 필요에 꼭 맞는 고유한 앱을 직접 만들 수 있고, 앱 제작 비용은 크게 낮아졌습니다.
로컬 AI가 다른 기술들이 걸어온 길을 따른다면, 하드웨어는 더 강력하고 저렴해질 가능성이 높고, 앱은 더 좋아질 가능성이 높으며, AI 코딩 도구는 앱 개발을 더 빠르게 할 가능성이 높습니다. 필자는 미래가 하이브리드일 것으로 예상합니다. 프라이버시, 비용, 편의성이 정당화하는 영역에서는 점점 더 많은 AI가 사용자 자신의 기기에서 실행되고, 나머지는 클라우드에 남을 것입니다.
아직은 때가 아닙니다. 오늘날 닭은 몇 마리 안 되고, 달걀은 그보다 더 적습니다.
필자의 관점. 이 보고서는 로컬 AI 시장에 대한 필자의 인상입니다. PromptQuorum를 개발한 경험, 필자가 직접 테스트한 앱 아홉 개, 인용한 공개 수치에 바탕을 두고 있습니다. 하드웨어와 앱이 모두 아직 빈약하며 둘 다 나아지리라는 결론은 필자 자신의 판단일 뿐, 출처가 내놓은 전망이 아닙니다.
핵심 요약. 닭은 몇 마리 안 되고 달걀은 더 적습니다. 오늘날 로컬 AI를 돌릴 수 있는 하드웨어는 적고, 써볼 만한 앱도 적으며, 저비용, 프라이버시, 최첨단 품질이라는 약속은 일반 사용자에게 아직 현실이 되지 못했습니다. 그러나 모델, 하드웨어, 앱 모두 빠르게 발전하고 있어 격차는 줄어들고 있습니다. 아직은 때가 아닐 뿐입니다.
전체 증거 및 부록모든 방법 설명, 계산, 가정 상자, 표, 주장 목록과 번호가 붙은 모든 출처를 담았습니다. 다운로드 없이 이 페이지에서 바로 열립니다. 전체 증거 파일(주장, 계산, 출처)은 영어로만 제공됩니다.
0. How to read this report
Every statement carries one of the labels below. A statement without a label is connective text and contains no factual claim. Where a number comes from a secondary report rather than the original publisher, the source list says so. Sources older than twelve months are marked older than 1 year in Section 11.
FACT Reported by a named third-party source (citation given).
AUTHOR-MEASURED Measured or observed by the author in PromptQuorum hands-on tests; evidence labels from the tests are kept.
CALCULATION Arithmetic performed on cited figures; the calculation is shown.
THIRD-PARTY FORECAST A third-party projection, not an observation.
ASSUMPTION An inference. The box names who makes it and what it rests on.
UNVERIFIED Stated by the author; no supporting data found yet.
Disclosure and method limits
The author founded PromptQuorum, which operates a content site, a directory of local-AI applications and a desktop application for multi-model prompting. Applications in the directory receive traffic from the site. No payment, sponsorship or vendor contact is recorded in the test notes.
The hands-on sample is nine tests by one tester, eight on one 24 GB Apple Mac (recorded as M3 Pro in earlier records; the LU record corrects its own “M3” note to M4, so the chip is recorded inconsistently) and one on an Android phone (12 GB). It is a convenience sample, not a random sample of the app market. The tester founded PromptQuorum, which lists the tested apps; for LU the developer supplied a founder statement and exchanged messages with the tester before and after the test (disclosure in the test record). No payment is recorded.
Several test records list the tester's own errors (for example, choosing the wrong model type, or judging features that were visible but not evaluated). These are kept in the underlying test pages and are summarised here where they affect a finding.
Consumer-spending data in this report is almost entirely US-only. No comparable EU household data was found.
1. Summary of findings
The report tests one hypothesis: that local AI adoption is held back by a small installed base of capable machines (the chickens) and a thin supply of applications that justify using them (the eggs). The table gives the result per component, with the strength of evidence.
#
Finding
Evidence
Strength
1
Capable hardware is a minority of even the best-case population. In the July 2026 Steam survey, 51.4% of opted-in PC gamers had 10 GB VRAM or more, 5.4% had 24 GB.[S1]
FACT
Medium: gamer sample, opt-in
2
Memory budget is consumed by the system before any model loads. A 24 GB Mac used 9.67 GB right after startup and 12.92 GB with ordinary apps open.[S22]
AUTHOR-MEASURED
Medium: one machine
3
Entry prices rose in 2026 while the usual pattern is falling prices. NVIDIA DGX Spark (128 GB): $3,999 at launch, $4,699 (Feb), $6,950 (Oct). 16 Gb DRAM: +49% from Q1 to Q2 2026.[S3, S4, S6]
FACTCALCULATION
High for Spark (Feb), medium for Oct figure
4
Few people pay. In the US, the most advanced market, 2.2% to 4.5% of households or consumers pay for a chatbot subscription (PNC 2.2%, Bank of America about 3%, YipitData/a16z 4.5%). No comparable EU data was found. Why it matters: cost saving is the main argument for local AI, and it only works if people spend money on AI today.[S11, S12, S13]
FACT
High: three independent card datasets
5
Result of the author's own hands-on tests (9 apps; eight on a 24 GB Mac, one on an Android phone): 4 of 8 scored apps rate 2 or below on core function. Test pages are linked in Section 4.2. This is the author's test series, not a market sample.[S22]
AUTHOR-MEASURED
Low-medium: n=9, one tester
5a
User interfaces are a recurring problem in the tests: six of seven scored apps rate 3 or below on onboarding (only LU rates 4); the same patterns (no model size or hardware fit, no remaining download time, unsuitable models offered without a memory warning, state not explained) recur across vendors. Detail in Section 4.2.[S22]
AUTHOR-MEASURED
Low-medium: n=9, one tester
6
The "6-12 months behind" statement is correct only for models small enough for a single consumer GPU (24-32 GB VRAM). The largest open models lag by ~4 months but do not run on consumer hardware. The author regards the lag figure as a marketing argument: the models it measures are not the ones users can run.[S14, S15]
FACTASSUMPTION
High (Epoch AI); assessment is the author's
7
Supply is moving in the direction of the thesis: Ollama raised $65M in July 2026; Gartner forecasts AI PCs at 55% of 2026 shipments, though the definition is NPU-based, not LLM-capable.[S8, S17]
FACTTHIRD-PARTY FORECAST
Medium
8
Speed: open models gain capability per parameter fast (density doubling about every 3.5 months to Apr 2025), new app releases are up 60-104% year on year, and memory supply is forecast to expand from 2027-2028 (forecasts disagree). Author's view: if local AI follows the general technology path, hardware and apps are likely to improve over the coming years; today not there yet.[S37, S38, S41]
FACTTHIRD-PARTY FORECASTASSUMPTION
Medium: trends and forecasts
Key takeaway. The data supports the "few chickens" half of the thesis for consumer hardware, and a qualified version of the "few eggs" half. It does not yet support statements about search interest, country-level adoption or app counts, which are listed as open items.
2. Thesis and definitions
2.1 The thesis
The classic chicken-and-egg problem asks which comes first. For local AI the problem is different: both sides are scarce at the same time. Few computers can run capable local models, and few applications exist that give a non-technical user a reason to buy such a computer.
2.2 Who the report is about
The reference user is a non-technical user who expects to download an app, understand what it does, optionally add a local model through Ollama or a direct download, and use it without a terminal, Python, or Linux. Developers and enterprise IT, who are today's main local-AI users, are discussed only as a contrast.
ASSUMPTIONMade by: Author (Hans Kuepper)
Enterprises and developers are the early market for local AI hardware; private end users follow later.
Based on: Author's market observation. Supporting data in this report: coverage describes Ollama as a developer platform, and DGX Spark is positioned as a developer system. No consumer-versus-enterprise split of local-AI users was found.
2.3 Definitions used
Term
Meaning in this report
Local AI
A model runs on the user's own device. Hybrid products that send some requests to a cloud service are noted as hybrid.
Open-weight model
A model whose weights can be downloaded. Not necessarily open source in the licence sense.
Frontier model
The most capable closed or open model at a given date, as measured by a stated index (here: Epoch Capabilities Index).
Capable machine
Not defined by a single threshold in sources. The report uses measured memory data from the author's tests and states the model size it applies to.
AI PC
A laptop or desktop with a chip for on-device AI (an NPU). Definitions differ: Canalys counts any chipset for dedicated AI workloads; Microsoft’s Copilot+ PC requires an NPU of 40 TOPS or more, 16 GB RAM and 256 GB storage.[S44, S45] This does not mean it can run a large language model well.
3. Few chickens: the hardware base
3.1 Best available data: the Steam Hardware Survey
Two different memories. A computer has system memory (RAM), which every program shares, and a graphics card has its own graphics memory (VRAM). A local AI model has to fit into one place to run at a usable speed, on a PC mostly the graphics card. This report therefore counts graphics memory (VRAM) only. Main-memory sizes are left out because they do not decide whether a PC can run a model: a PC can have 32 GB of RAM and 8 GB of VRAM. Macs and some other designs share one pool, so for them the single memory figure is the one that counts.
No public dataset reports the share of all personal computers that can run a given local model. The closest public proxy is Valve's monthly Steam survey. It is opt-in, covers gamers, and is therefore likely to overstate capability compared with the general PC population.
Steam survey, July 2026: share of surveyed systems by GPU memory (VRAM)
Source: Steam July 2026 as reported by tBreak, Hardware Busters, Club386. Listed buckets total 74.0%; the remainder is other sizes. 10 GB or more: 51.42% (tBreak). The default Steam chart combines Windows, macOS and Linux; on Windows only, 8 GB remains the most common at 33.20%. September 2026: 16 GB 27.21%, 8 GB 26.71%.
FACT The July 2026 survey shows 16 GB VRAM as the largest single bucket for the first time (25.90%).[S1]FACT Only 5.43% of respondents had 24 GB of VRAM, and a 32 GB bucket was not found in the retrieved reports. FACT macOS accounted for 1.92% of Steam systems in September 2026, so the survey says little about Mac users.[S2]
Key takeaway. Among opted-in gamers, the 16-32 GB range is now common, but 24 GB or more of VRAM is a minority. For the general population of laptops and PCs no comparable data exists, and the true share is probably lower (assumption above).
3.2 The system uses memory before the model does
A model size that fits in total memory does not fit in free memory. The author's tests give measured values on a 24 GB Apple Mac.
24 GB Mac: memory in use vs. memory a 17 GB image model needs
Source: author's Unsloth test, Activity Monitor readings (measured) and the app's own memory message (vendor-stated need). "GB left" = 24 minus used (calculation). Readings were taken at different moments in the test.
AUTHOR-MEASURED Activity Monitor showed 9.67 GB used right after startup and 12.92 GB used in a first capture with a browser, office and chat applications open.[S22]AUTHOR-MEASURED The Unsloth app refused a 17 GB image model; it counted about 2 GB usable (7 GB free) before a reboot and about 6 GB usable (11 GB free) after.[S22]CALCULATION 24 − 9.67 = 14.33 GB left after startup, which is below the 17 GB the app stated for that model.
AUTHOR-MEASURED A 27B-parameter model in 4-bit format ran in chat at 11.0 tokens per second on the same machine, but reading a web page and an event search both ended in "Ran out of memory" (at 34.4k context).[S22]
ASSUMPTIONMade by: Author (Hans Kuepper)
Even 32 GB of VRAM may not be enough for a normal user who keeps an editor and other applications open next to the model.
Based on: Author's reasoning from the 24 GB measurements above. No test on a 32 GB machine exists in the data; the statement is an extrapolation.
AUTHOR-MEASURED Guides often say macOS uses about 4 GB. In the author’s normal use of a 24 GB Mac the system used 9.67 GB right after startup. The guide figure describes a bare system; the measured figure is what an ordinary user experiences, and it is the one that decides how much memory is left for a model. The figures are from a normal working machine, not a laboratory set-up.[S22]
3.3 The price of a capable machine
NVIDIA DGX Spark (128 GB), list price in USD
Sources: NVIDIA forum (Feb 2026, primary); Implicator.ai (Oct 2026, secondary, NVIDIA page not retrieved). Partner GB10 systems were reported at $7,000-9,000 (Tom's Hardware via Implicator.ai).
CALCULATION $6,950 / $3,999 = 1.74, i.e. +74% since launch. The February increase was +$700 (+18%); the October increase was +$2,251 (+48%). NVIDIA attributed both to memory supply constraints.[S3, S4]
FACT Component pricing shows the same direction: the average 16 Gb DRAM price rose from $19.2 to $28.5 between Q1 and Q2 2026.[S6]CALCULATION 28.5 / 19.2 = 1.48, i.e. +48% to +49%. One outlet reports +89% for the same data; the underlying figures do not support it.[S6]
FACT IDC expects the DRAM/NAND shortage to last into 2027 and warns of up to 8.9% fewer PC shipments in the worst case; Dell and Lenovo announced 15-20% price increases.[S7]
FACT Counter-example: Apple raised entry prices of the M5 MacBook Air ($100) and the 14-inch MacBook Pro with M5 Pro ($200) in March 2026 but left RAM-upgrade pricing unchanged (24 to 48 GB on the 14-inch: $400, as in the previous generation).[S5] Later reports of further Apple price rises exist but come from a single observer and were not verified.
Key takeaway. The usual pattern of falling prices for new hardware did not hold in 2026 for memory-heavy machines. The exception is Apple, whose RAM upgrade pricing was unchanged at the March 2026 launch.
3.4 The installed base: forecast versus capability
Gartner forecast: AI PCs as share of worldwide PC shipments
Gartner, August 2025 (older than 1 year). Shipments: 77.8M in 2025, 143M forecast for 2026. 2024: about 38M (IT Reseller, citing Gartner). Forecast pre-dates the 2026 memory shortage.
CALCULATION 77.8M + 143M = 220.8M AI PCs shipped in 2025-2026 on Gartner's forecast; with about 38M in 2024 the three-year sum is about 259M. This is consistent with the statement that a base of "hundreds of millions" of AI-labelled devices will exist well before 2030.
New laptops and PCs: AI-enabled share, and what the label means
Source
Period
AI-enabled share of shipments
Definition used
Date flag
Canalys
Q2 2024
14% (8.8M units)
Desktop or notebook with a chipset for dedicated AI workloads, such as an NPU[S44]
Not formally defined on the retrieved page; tied to on-device AI features and the processor platform[S8]
older than 1 year
Gartner (forecast)
2026
54.7% (143.1M units)
Same; forecast made in August 2025, before the 2026 memory shortage[S8]
older than 1 year
IDC / Omdia
Q3 2026
Total shipments -20.1% / -21.2% year on year (62.7M / 58.1M units); no AI share given[S43]
-
current
CALCULATION IDC: 62.7M / 78.5M = 0.80, i.e. -20.1%. Gartner 2026: 143.1M AI PCs at 54.7% implies total 2026 shipments of about 262M, which was the forecast before the Q3 fall.
FACT Microsoft’s Copilot+ PC floor of 16 GB RAM is the only memory requirement found in any AI-PC definition; none is based on running a large language model.[S45]
4. Even fewer eggs: applications and willingness to pay
4.1 How many people pay for AI at all
The author's starting point was that about 2-3% of US households pay for AI, so a "free" local alternative has limited appeal. Three independent US card datasets confirm a low single-digit share. This report uses only that range.
US: share paying for AI, by source
PNC and BofA: payment data of their own customers. a16z/YipitData: card panel, paid personal ChatGPT, Gemini or Claude subscriptions only (up from 2.1% a year earlier). The datasets use different bases (households vs. consumers), so the range 2.2-4.5% is reported rather than one number.
FACT Card data from PNC (2.2%, May 2026), Bank of America (~3%) and YipitData via a16z (4.5%, up from 2.1% a year earlier) agree on a low single-digit share.[S11, S12, S13]FACT The top 1% of payers account for 19.5% of observed consumer AI spend, averaging $903 per month.[S13]
4.2 What the hands-on tests show
These are results of the author's own hands-on work, not a market survey. Nine applications were tested between 6 and 24 October 2026, eight on a 24 GB Apple Mac and one on a 12 GB Android phone. Each app name links to the full test page with evidence labels, scores and the tester's own errors.[S22]
*Score proposed from the evidence; the tester has to confirm. Core-function scores are the tester's dimension scores; the scale maximum is not stated in the data. n/a = not tested.
Core-function score per tested app (higher is better)
Axis set to 0-5 (scale maximum is not stated in the test data). LTX Desktop is not scored.
AUTHOR-MEASURED Each row lists interface problems from the test findings (finding IDs and evidence label as recorded; "observed" unless stated), the strengths recorded, and any misreading by the tester that touches the same point. Interface problems are not scored separately in the tests; the table is a reading of the findings.[S22]
Model list: no size, task or hardware fit (DT-01, High). Welcome dialog suggests FLUX.1 with no memory caveat; downloads read as 25-57 GB on a 24 GB Mac (DT-02, High). No guard against a video-restoration model for a text prompt, result a gray square (DT-03, High). Download shows no remaining time (DT-04); app unusable during download (DT-05). Controls unclear (DT-11, tester view).
Install and image speed once a suitable model ran (scores 4).
Wrote that the list has no search (a search field exists); judged iteration without evaluating Version History or Edit tools.
Model cards: no size, task or fit (DB-01, High). FLUX cards say the app must be updated although newly installed (DB-02, High). Percentage only, no size or time (DB-04); a timed-out download offered only 'Download Again' (DB-05). Input Strength 100 crashes the app, per notes without screenshot (DB-18, High). App does not explain what a model is (DB-21, tester view).
Queue for several pictures; 'Actions' button on results (DB-08, DB-24).
Tester suspects he mishandled Input Strength (assumption).
Screen-recording permission step unclear (BB-06). In-app assistant did not name the macOS setting after two questions (BB-07, High). No remaining time (BB-05). Announced 11 GB, downloaded 4.9 and 5.7 GB, unexplained (BB-04). App restarted when deleted (BB-10, High).
Clear three-model dialog with RAM needs (BB-03); 312 MB install, one approval (BB-01).
Wrote there is no clear way to set up screen access; an 'Open System Settings' link is shown.
Toolbar icon opens a full-screen page, sidebar found by right-click only after help from ChatGPT (PA-04, High). Welcome-video shortcut did not work (PA-05). First side-panel question showed only a cursor (PA-08). Two chats returned empty 'Thought for 0ms' replies (PA-07, High). 'Add MCP server' does nothing visible (PA-15).
Word count of a page worked (PA-09).
Missed the store link in the README; sidebar route miss is shared with the welcome video.
First screens do not say what the app is for (MS-01, tester view). Image-only and cloud models sit in the same picker as chat models (MS-05). File upload looped between knowledge base and chat (MS-08, High, tester view). No local image model selectable in Media Studio (MS-10, High).
Existing Ollama models found at once (MS-03); hardware-fit hint on model cards (MS-11).
Believed the Ollama connection had failed; it had worked. Wrote the app does not say 'Ollama'; a screenshot shows the header.
Picker shows parameter counts, not gigabytes (UN-04). Model start shows a spinner with no percentage or time for two to three minutes or more (UN-23). 13-step dependency install of 4-5 minutes with no estimate (UN-02).
Hardware readout, download time left, size warnings (UN-09, UN-11); refusal message explains cause and names an override (UN-20); context meter and tokens per second (UN-24).
Read parameter counts as sizes; wrote there is no token counter (one is shown).
Setup listed 129.1 GB of downloads (five files). The main model, LTX 2.5, is 41.5 GB, larger than the Mac's 24 GB of memory, and no memory warning was visible in the screenshots (LTX-03, High, observed). Download size is not runtime memory (the test record's own assumption); the vendor README lists 32 GiB of RAM as the floor for local generation (LTX-01, vendor claim). The 129.1 GB total also exceeds the 32.9 GB of free disk; whether the app blocks was not tested (LTX-04). Hugging Face sign-in required (LTX-05).
PDF (872 KB) and Excel file attached with the paperclip not read; Word file listed as 20 chunks yet reported unreadable (LU-10, LU-11, LU-13; High for PDF and Word). Download dialogs show percentage and MB/s but no time left (LU-15). Sampling settings (Temperature, Top P, Top K) unexplained (LU-23, tester view). Troubleshoot window would not close (LU-24).
Three-step setup, existing models found (LU-04); size, memory need and fit label on image-model cards (LU-17); context meter (LU-09); built-in explainer (LU-20); coding agent built a working game (LU-29); A/B compare and benchmark (LU-27, LU-28).
Expected the engine tick to toggle; tested .xlsx, which the Docs zone does not list; judged image quality on one 512 px starter model; did not record whether the embedding model was installed.
Never says which model it downloaded (CAE-10). Web-search button has two blue shades, no label (CAE-06). App does not stress that the AI runs on the phone (CAE-14, tester view). Unclear when the web interface is on (CAE-15).
Download with progress bar, size and ETA (CAE-01); edit, regenerate, copy, read-aloud buttons (CAE-16).
-
Onboarding and guidance scores per app (higher is better)
Source: dimension scores in the author's hands-on tests (S22). Caelum and LTX Desktop were not scored on these dimensions; LU is scored. Axis 0-5 (scale maximum is not stated in the test data).
Draw Things, DiffusionBee (2); Unsloth shows parameter counts only (partial)
BoBe (RAM needs), Msty (hardware-fit hint), LTX Desktop (sizes in setup), LU (size, memory need, fit label)
Page Assist, Caelum
Unsuitable model or download offered without a visible memory warning
Draw Things (FLUX.1 suggestion, download sizes above 24 GB), LTX Desktop (41.5 GB main model on a 24 GB Mac) (2)
BoBe (RAM stated per model), Unsloth (several warnings), LU ('Too big for this Mac' labels)
others
App does not explain its own state or the model used
BoBe (assistant), Page Assist (empty replies), Caelum (model unnamed), DiffusionBee (what a model is)
Unsloth (refusal explains cause)
others
ASSUMPTIONMade by: Author (Hans Kuepper)
Very few local-AI apps have a user interface in which everything is developed, and for non-technical users this is now a bottleneck as large as hardware.
Based on: The tests above: six of seven scored apps rated 3 or below on onboarding (only LU rated 4) and only two rated 3 on guidance; the same four patterns recur across different vendors. Limits: n=9, one tester, convenience sample; at least five records list a related misreading by the tester, so counts are an upper bound; whether a missed control is an interface fault is a judgement. No data exists on how a typical non-technical user would fare. Counter-evidence: Unsloth's interface was rated polished by the tester, LU's first-run setup and in-app explainers were rated strong, and BoBe, Msty and Caelum have clear elements recorded.
Key takeaway. The tests support a pattern (missing model information, missing time estimates in some apps, gaps in editing) but not precise percentages. With nine tests the report can state counts, not rates.
4.3 The size and shape of the app market
No app store for local-AI software exists; applications are spread across GitHub, vendor websites and the mobile stores. The author's own directory gives one structured view.
PromptQuorum directory: listings per category group, 23 Oct 2026
Source: directory category counts (author's own directory). Group counts may include an app in more than one group; the sum of 315 is therefore an upper bound on distinct apps. General chat clients: 74 listings (23% of 315). Contract review: 1 listing. Curated directory: not a census of the market.
AUTHOR-MEASURED Within the directory, 74 of up to 315 listings are general chat clients, which is consistent with the observation that many local applications are chat front ends on top of a downloaded model.[S23]
Store searches, 23 October 2026
Store
Query
Result
Observed
Apple App Store (apple.com/de)
local AI
10 results
9 are local-AI apps by title: Locally AI (LM Studio), PocketPal AI, Liquid Apollo, Google AI Edge Gallery, Enclave, Oscilla, Atomic Chat, pocket, Draw Things. One (Grok) is not a local app.[S34]
Google Play (Germany)
local ai
29 listings on the page as captured (16 PDF pages)
Titles such as LocalAI, Ollama Local AI, PocketPal AI, Off Grid AI, Private Mind, Gemma 4 Offline, Llamatik, Caelum, Secret AI. At least 4 listings name a connection to Ollama, vLLM, OpenRouter or LM Studio in their title or screenshot, so they may be clients for a server rather than on-device models. Several carry a rating of 2.2-4.2 stars; many carry none.[S35]
FACT One documented search per store shows 10 results (Apple) and 29 listings (Google Play) for a local-AI query. FACT The author's statement "about ten iOS apps" matches the first Apple result page.
FACT For scale: the Apple App Store lists about 2.75 million apps and Google Play about 2.65 million (aggregator, 23 Oct 2026).[S36] Against that, a local-AI query surfaces tens of apps. ASSUMPTIONAuthor: it is very early in the cycle and there is basically nothing available yet for ordinary users.
ASSUMPTIONMade by: Author (Hans Kuepper)
Much of the available local-AI software is distributed through GitHub, and GitHub is not a place for ordinary users, so for them it is effectively not available.
Based on: Test records (S22): GitHub is the vendor page, README or release source in five of the first eight tests (Page Assist, LTX Desktop, Unsloth, DiffusionBee, BoBe); LU's source was not checked. Page Assist's README leads with a source build that needs a terminal and Bun; the tester's first attempt failed on a PATH error (a Bun matter, not Page Assist itself) and he missed the store link lower on the page. GitHub topic pages list 5,794 public repositories for 'ollama' and 9,477 for 'local-llm' (S46); counts include libraries and forks, overlap between topics not removed. The author describes the field as thousands of forked projects where nobody knows what is current (author's observation, not measured). Limits: five test records, one tester.
UNVERIFIED
These are search result pages, not store-wide counts. The result number depends on the store's ranking and on how many results load on one page, so Apple 10 versus Google Play 29 is not a measure of which store has more local-AI apps. Titles are matched against the query, which inflates the number of listings named 'Local AI'. App function (on-device or server client, chat or other) was judged from titles and screenshots only; no app was installed.
UNVERIFIED
"Most iPhone apps are chatbot extensions." Not verified. Of the nine Apple results, one (Draw Things) is an image generator per the author's own test; the function of the others was not checked beyond their titles. The directory figure (74 of up to 315 listings are chat clients, all platforms) is the only count available.
FACT Apple's WWDC 2026 announcements added image input to its Foundation Models framework and a Core AI framework for running other local models on Apple silicon.[S24] This provides on-device model access to third-party apps without the user installing a separate runtime.
4.4 Capital and corporate involvement
Runtime and platform layer
Company / segment
Reported funding or scale
Source quality
Ollama
$65M Series B, 9 Jul 2026 (Theory Ventures; Benchmark, 8VC, Y Combinator and others); $88M total; 14 employees; 8.9M monthly active developers (company-stated)
Multiple outlets; user number is company-stated
LM Studio
~$19.3M over two rounds (May 2025)
Aggregator; unverified
Platform owners
Apple (Foundation Models, Core AI), Microsoft (Copilot+ PCs), NVIDIA (DGX Spark)
Application layer (apps for end users): what the search found
App / company
Funding trace found
Source quality
GPT4All (Nomic AI)
$17M Series A, July 2023, led by Coatue (Y Combinator, Spark Capital and others). Nomic also sells embedding and data-mapping products, so the round is not purely for the app. older than 1 year
Aggregator page (startupintros)
AnythingLLM (Mintplex Labs)
$0.5M convertible note; investors Y Combinator, UpHonest Capital, Pioneer Fund. CB Insights lists the note at 1 Jan 2022, which pre-dates the product; date unverified. A second round is behind a paywall.
No funding trace found. Searches returned app-store pages, reviews and unrelated companies. Venture rounds are normally announced publicly; the author reads the absence as evidence of little investment (assumption below).
Result.FACT Funding is visible at the runtime layer (Ollama, LM Studio). At the consumer-application layer one older round (Nomic, 2023) and one small note (Mintplex) were found, and no trace for five other named apps; six more apps were not searched.[S17, S18, S27, S28]AUTHOR-MEASUREDAuthor: one app he is in contact with has just found an investor; app and investor not named, not verified here.
ASSUMPTIONMade by: Author (Hans Kuepper)
Very few investments are made in consumer local-AI applications. The market is early.
Based on: Venture investments are normally made public, and the search found none for five of eight named apps (LU, the ninth test, was not searched) and only one older and one small round for the other two. Limits: web search only; no paid database (Crunchbase, PitchBook) was queried; angel and unannounced rounds would not appear. The newly reported investor in one of the author's contact apps is one data point against a zero count and is not verified.
FACT A figure of "52 million monthly downloads" for Ollama circulates on secondary sites and conflicts with the company-stated 8.9 million monthly active developers; it is not used here.
4.5 Supply of applications: the vibe-coding argument
ASSUMPTIONMade by: Author (Hans Kuepper)
Within one to two years, hundreds of thousands of apps will exist, including local-AI apps, because vibe coding lowers the cost of building software.
Based on: Author's direct experience building PromptQuorum applications with AI coding tools, and press reports about software companies being replaced by in-house builds. No time series of local-AI app counts was found to test the forecast.
FACT A figure of $285 billion in software-stock value lost in 48 hours in early February 2026 is attributed to vibe-coding fears in one derivative article; the original could not be verified.[S25] It is not used as evidence in this report.
FACT Hugging Face reports that 1.5% of repositories account for 99.2% of downloads and that 85.6% of models have fewer than 200 lifetime downloads.[S19]ASSUMPTIONInference from that distribution: a large number of new apps does not by itself imply a large number of used apps; usage in open ecosystems is highly concentrated.
Key takeaway. The application layer is real but small and fragmented. Funding exists for runtimes, not clearly for consumer applications. Claims about the number of apps need a counted series, which the quarterly format can provide.
5. Local models versus frontier models
A frequent claim is that local models are six to twelve months behind the frontier, or only about four months for the largest. The author regards this as a marketing argument: it is offered as a reason to switch, and it does not describe what an ordinary machine can run. The source for the figures, Epoch AI, defines them for specific model classes.
Epoch AI: lag of open-weight models behind closed frontier (months)
Epoch Capabilities Index (ECI) for the first two rows; benchmark comparisons (GPQA-Diamond, MMLU-Pro, AA Intelligence, LMArena) for the consumer-GPU row, defined as up to 28B parameters (RTX 4090) or 40B (RTX 5090). The consumer-GPU analysis is older than 1 year. Epoch cautions that real-world usefulness lag is likely longer than benchmark lag because the best open model per benchmark is more likely to be over-fitted.
FACT Since January 2026 the most capable open-weight models lag the closed frontier by about four months; the earlier figure was three months.[S14, S16]
FACT Models small enough for one consumer GPU reach frontier-level benchmark scores with a lag of 6-12 months.[S15]
FACT Hugging Face reports that Chinese labs released the largest open model almost every month of 2026, with ceilings between 754 billion and 2.78 trillion parameters.[S19]
CALCULATION A model of 754 billion parameters stored in 4-bit format needs more than 350 GB for weights alone (754 x 0.5 bytes), well above any consumer memory in Section 3. The 4-month figure therefore describes models few users can run.
What the two lag figures actually cover
Model class
Parameters
Memory for weights (4-bit, calculated)
Hardware needed
Measured lag behind frontier
Source
Largest open models
754 billion to 2.78 trillion (Chinese labs' monthly largest, 2026)
about 377 GB to 1,390 GB
Data centre or multi-machine cluster
about 4 months (Epoch Capabilities Index, since Jan 2026)
S14, S19
Models for one consumer GPU
up to 28B (RTX 4090) or 40B (RTX 5090)
about 14 GB / 20 GB
GPU with 24 GB / 32 GB memory
6-12 months (benchmarks; older than 1 year)
S15
Models for a typical laptop or mid-range GPU
about 7-8B
about 4 GB
8-16 GB total memory
No figure found
-
CALCULATION Weights only: parameters x 0.5 byte at 4-bit quantisation. Excludes context cache, runtime overhead and the operating system. Section 3.2 shows the system and open apps already use about 10-13 GB of a 24 GB Mac.
In plain terms. Two different lags are quoted and they describe different things. "About 4 months" compares the best open model with the best closed model; the best open models are too large for any consumer machine. "6-12 months" compares models small enough for one consumer GPU with 24-32 GB of memory with the frontier. Neither figure covers a computer with less memory.
FACT Epoch adds that the best-scoring open model on a benchmark is more likely to be over-fitted to it, so the lag in everyday usefulness is probably longer than the benchmark lag.[S15]
ASSUMPTIONMade by: Author (Hans Kuepper)
The "local models are only months behind" statement is a marketing argument and cannot be used in practice by the reference user.
Based on: Table above: the 4-month figure covers models needing about 377 GB or more; the 6-12 month figure covers models that need a 24-32 GB GPU and is older than 1 year; no figure exists for a typical laptop or phone. Tests (S22): a 27B model on a 24 GB Mac ran out of memory on web-page reading and an event search; Caelum, a 1 GB model on a phone, gave a false news answer and a logic answer rated 5/25 by ChatGPT as judge (judge not independent, not verified). Limits: few tests, one tester, no side-by-side local-versus-cloud benchmark on the same tasks.
Key takeaway. The lag figures are real on benchmarks, but they describe models users cannot run (4 months) or older consumer-GPU data (6-12 months, older than 1 year). In the author’s view they are a marketing argument, because for a typical laptop or phone no figure exists and the tests show failures.
6. Demand signals and what is changing
6.1 Developer-side adoption is rising
Hugging Face downloads in 2026, by model family (billions)
Sources: Hugging Face report as cited by Bloomberg and a French summary. Alibaba itself claims more than 3 billion in six months; Hugging Face's own count is lower. Periods are not identical (Google and Meta: 2026 to date; Qwen: Jan-Aug 2026). Download counts are not user counts.
FACT Models under 1 billion parameters account for 83% of all-time Hugging Face downloads.[S19]FACT Ollama's developer base doubled between January and July 2026 and the company reports about one million new installs per week (company-stated).[S17]
6.2 Search interest
Google Trends, worldwide, weekly: "local ai" and "local llm" (index 0-100)
Source: Google Trends CSV supplied by the author (S29). Index is relative: 100 = highest weekly value of either term in the window. Only 12 months are available, so seasonality cannot be seen.
AUTHOR-MEASURED"local ai": about 31 in the first four weeks (Oct-Nov 2025); 100 in the week of 31 May 2026 (about 3.3 times the Oct 2025 level); 15 in the week of 18 Oct 2026, half the Oct 2025 level. Average of the last four weeks 19.8, which is 36% below the first four weeks (30.8).[S29]
AUTHOR-MEASURED"local llm": 6 in Oct 2025; peak 28 (week of 7 Jun 2026); 5 in the latest week. Roughly one quarter of the scale of "local ai" throughout.[S29]
AUTHOR-MEASUREDThe fall is abrupt, not gradual. "local ai" went from 72 to 41 between the weeks of 5 Jul and 12 Jul (-43%); "local llm" from 21 to 10 (-52%) in the same week. Both had been flat at about 85 and 25 for the previous four weeks.[S29]
AUTHOR-MEASURED From August to mid-September both terms sat at roughly their Oct 2025 levels (local ai 25-38), apart from one week at 49 (20 Sep), then fell again to 15.[S29]
UNVERIFIED
The divergence cannot be explained. Google Trends shows local-AI search interest falling from June and stepping down in July 2026, while other indicators rise: the Ollama developer base doubled between January and July (company-stated), Hugging Face downloads are high, and the author reports that other trends keep rising. The developer indicators available here end in July, so the August-October period cannot be compared. Both search terms moving by similar percentages in the same week fits a change in the data or in what is counted as well as a real change in interest; a search for a documented Google Trends methodology change in July 2026 found nothing. The export covers 12 months only, so a seasonal pattern cannot be excluded. No cause is claimed.
6.3 Privacy as a driver
FACT In a DuckDuckGo survey (n=1,944 US adults; the sponsor sells a privacy product), 32% of chatbot users had shared information they withheld from friends, relatives, colleagues or medical professionals; 53% did not know or were unsure that conversations may be used for AI training.[S20]FACT In a smaller academic survey (n=300), 63.6% reported privacy concerns about conversations with ChatGPT.[S21]
ASSUMPTIONMade by: Author (Hans Kuepper)
Privacy is a weak sales argument for local AI because users have accepted data-for-service trades with Google and Meta for twenty years.
Based on: Author's observation of consumer behaviour. The surveys above show high stated concern together with high disclosure, which fits the author's view of behaviour, but they also show low awareness, which points to an information gap rather than to indifference. The data does not decide between the two.
6.4 Geography
UNVERIFIED
"China is a big local-AI market; other countries are underdeveloped; awareness is minuscule." No country-level data on local-AI users was found. What exists describes model supply (Chinese labs lead in the size of open releases and in Qwen downloads, and Hugging Face attributes 41% of platform downloads to Chinese models in the prior year) and does not measure where local AI is used.
Review of the "wave theory"
FACT The general concept is documented. Comin and Hobijn find that most technologies are introduced in advanced economies and then "trickle down" to lagging countries, and that diffusion sped up after World War II.[S31]older than 1 year
FACT For generative AI, adoption is uneven by country. Microsoft reports 17.8% of the world's working-age population using AI in Q1 2026, 27.5% in the Global North against 15.4% in the Global South, and a range from 70.1% in the UAE to 31.3% in the US (rank 21, up from 24).[S30]
Part of the theory
Evidence
Result
Technology adoption is uneven across countries and reaches some later
Comin and Hobijn (S31); Microsoft 2026 country data (S30)
Confirmed
New technology starts in advanced economies
Comin and Hobijn (S31); North-South gap in S30
Confirmed
The US specifically leads each wave
Supported for where AI models are built; not for usage: US ranks 21st in Microsoft's AI-use share, behind the UAE and others (S30)
Narrowed
Local AI follows the same wave, and non-English markets lag
No country-level data on local-AI use found; Google Trends (S29) is worldwide only
Unverified
ASSUMPTIONMade by: Author (Hans Kuepper)
Local AI spreads country by country in the same pattern as other technologies.
Based on: The general concept is backed by S30 and S31, so the theory is not "mad" and is kept. The step to local AI is an extrapolation without data. The "US leads" part is replaced by "advanced economies lead", because the AI-use ranking does not put the US first.
7. Trade and supply-chain implications (EU, China)
FACT The memory shortage is global and driven by allocation to AI data-centre products; it affects EU buyers through higher PC and component prices.[S6, S7]
UNVERIFIED A user post on NVIDIA's forum reports an EU marketplace price of 4,800 EUR for the 128 GB DGX Spark in February 2026 against $4,699 in the US. Currency, VAT treatment and date were not verified.[S26]
FACT The largest open-weight models and the most downloaded model family are published by Chinese labs; the Qwen family is described as Apache-2.0 licensed on one secondary source.[S19] For EU users this makes model supply dependent on Chinese labs, while the hardware supply chain (memory, GPUs) is dependent on US-designed and Asian-manufactured parts.
UNVERIFIED EU regulatory effects (AI Act, data-protection rules) on local versus cloud deployment were not researched for this report.
8. Outlook, with assumptions labelled
Statement
Type
Who / basis
Memory prices stay elevated into 2027
THIRD-PARTY FORECAST
IDC (Dec 2025). One secondary report cites a view to 2028; unverified.
Future is hybrid; a growing share of AI moves local where privacy, cost or convenience justify it
ASSUMPTION
Author. No measured local share exists; needs a tracked indicator.
Local AI will grow faster in percent than the overall AI market
ASSUMPTION
Author. Plausible because it starts from a lower base; not tested.
Mobile will churn faster than PCs and Apple users will pay for on-device AI
ASSUMPTION
Author. Apple's WWDC 2026 changes are facts (4.3); the willingness to pay is not measured.
An on-device email organiser would succeed
OPINION
Author. No data; product opinion.
Projection: what could be ready in the coming years
Driver
What the sources say
Direction
Type
Model efficiency
Capability density of open models doubled about every 3.5 months, Feb 2023 to Apr 2025 (51 models). A model with half the parameters matches the best open model about every 3.5 months.[S37]
Smaller models become more capable
FACT for the period; extending it is ASSUMPTION
Memory supply
Counterpoint: new output from 2027, more capacity in 2028, prices reversing by 2028. TrendForce: consumer prices keep rising next year. IDC: shortage into 2027. Samsung warning of worse in 2027-2028 (headline only, unverified).[S38, S7, S39]
Forecasts disagree
THIRD-PARTY FORECAST
AI PCs
Gartner: 31% of shipments 2025, 54.7% 2026, 'the norm' by 2029. Definition is not about running large models; forecast pre-dates the memory shortage.[S8]
More devices with on-device AI chips
THIRD-PARTY FORECAST
Consumer GPUs
RTX 60 series rumoured for H1 2027 or 2028; one leak gives up to 32 GB for the top card, equal to the 32 GB class of the RTX 5090 in S15. Unconfirmed.[S40, S15]
Unclear
UNVERIFIED
App supply
New app releases +60% (Q1 2026 versus Q1 2025) and +104% (April to date), Appfigures; +30% in 2025 to nearly 600,000, Sensor Tower. AI coding suspected, not measured.[S41, S42]
More apps, quality unknown
FACT for counts; cause ASSUMPTION
ASSUMPTIONMade by: Author (Hans Kuepper)
If local AI follows the general path of technology, capable hardware is likely to become available at mass-market prices over the coming years, apps are likely to get better, and AI coding tools are likely to speed up app development. Local AI is developing at great speed but is not there yet.
Based on: The table above: model efficiency doubling about every 3.5 months (measured to April 2025), memory supply expansion forecast from 2027-2028 by Counterpoint, Gartner's AI PC trajectory, and the rise in new app releases. Limits: memory forecasts conflict; the efficiency trend is a fit on 26 months of data and extending it to three or four years is an extrapolation; no source projects the share of machines able to run a 30B model; the app data cover all apps, not local AI, and measure neither quality nor use.
Proposed indicators to track every quarter
For the series to be quotable, each report should repeat the same measurements so that changes can be read as a time series.
Indicator
Source
Q4 2026 value
Share of Steam systems with 16 GB+ VRAM; 24 GB+; 32 GB+
Replaced by US card-data range: 2.2-4.5% of households or consumers pay. Menlo 2025 (~3%) is an implied revenue ratio; Menlo 2026 reports 55% of AI users pay (survey, different denominator, not used in the report body).
Google Trends interest in local AI fell substantially since summer 2026
Confirmed
'local ai' peak 100 (31 May 2026), 15 on 18 Oct 2026; 36% below Oct-Nov 2025 on four-week averages; abrupt step in the week of 12 Jul. Unexplained: other indicators (Ollama developers, Hugging Face downloads, author's other trends) rise while Trends falls. Worldwide index, 12 months.
Runtime layer: funded (Ollama $65M, LM Studio ~$19M). Application layer: traces for 2 of 8 named apps, both small or old (GPT4All/Nomic $17M, 2023; AnythingLLM $0.5M note); none for 5. Author's inference: rounds are normally public, so few exist, early market. Author reports one contact app now has an investor (unverified, unnamed). No paid database queried.
Very few local-AI apps have a good, fully developed user interface
Partly supported
In the author's 9 tests one app (LU) scored above 3 on onboarding (7 scored); four interface patterns recur. n=9, one tester, tester misreadings in at least 5 records; no data on typical users.
Hardware and apps are likely to improve over the coming years if local AI follows the general technology path; AI coding tools speed up app development
Opinion / assumption
Supported directionally: efficiency trend (to Apr 2025), memory capacity forecast from 2027-2028 (forecasts conflict), new app releases +60-104%. No projection of machines able to run a 30B model; app data not specific to local AI.
Sources are listed with publication date. older than 1 year marks sources published before 23 October 2025. Secondary reports are named as such; where a primary page was not retrieved, this is stated.
S15Epoch AI Data Insight: models runnable on a single consumer GPU match frontier with a 6-12 month lag (<=28B params for RTX 4090, <=40B for RTX 5090). Publication date not shown in retrieved copy; text refers to mid-2026 as a future date, so written in 2025. 2025 (undated)older than 1 year https://epoch.ai/data-insights/consumer-gpu-model-gap
S21arXiv 2508.06760: privacy norms around LLM chatbots, n=300 respondents. Aug 2025older than 1 year https://arxiv.org/pdf/2508.06760
S22PromptQuorum hands-on tests (author's own), nine tests published 9-24 Oct 2026. Index of test pages: https://www.promptquorum.com/power-local-llm/ . Individual slugs: draw-things, bobe, page-assist, msty, unsloth, diffusionbee, ltx-desktop, caelum, locally-uncensored (suffix -hands-on-test). 9-21 Oct 2026 https://www.promptquorum.com/power-local-llm/unsloth-hands-on-test
S23PromptQuorum Local LLM Software Directory: category listing counts as of 23 Oct 2026 (author's own directory). 23 Oct 2026 https://www.promptquorum.com/
S28CB Insights: Mintplex Labs (AnythingLLM) $0.5M convertible note; investors Y Combinator, UpHonest Capital, Pioneer Fund (partial, second round paywalled). 2022 (round date as listed)older than 1 year https://www.cbinsights.com/company/mintplex-labs/financials
S29Google Trends export supplied by the author: terms 'local llm' and 'local ai', Worldwide, All categories, weekly, 19 Oct 2025 to 18 Oct 2026 (53 weeks). Index 0-100 relative to the highest point of the two terms. export 23 Oct 2026 https://trends.google.com/trends/explore
S31Comin and Hobijn, New York Fed Staff Report 169 (cross-country technology adoption): new technologies typically appear in advanced economies and 'trickle down' to lagging countries; diffusion sped up after World War II. Publication year not retrieved; the report series number places it well before 2026 (assumption). year not retrievedolder than 1 year https://www.newyorkfed.org/medialibrary/media/research/staff_reports/sr169.html
S32Ollama model library, qwen3-vl: described as the most powerful Qwen vision-language model; accepts image and video input. captured 23 Oct 2026 (page shows '5 months ago') https://registry.ollama.ai/library/qwen3-vl:2b
S33Ollama model library, qwen3-coder: listed input type is text only; sizes 30B (19 GB) and 480B (290 GB). captured 23 Oct 2026 https://ollama.com/library/qwen3-coder
S34Apple App Store search 'local AI' (apple.com/de), screenshot PDF supplied by the author, captured 23 Oct 2026 15:23: 10 results. 23 Oct 2026 https://www.apple.com/de/search/
S3642matters store statistics: 2,749,633 apps in the Apple App Store and 2,651,095 in Google Play (page dated 23 Oct 2026, aggregator). 23 Oct 2026 https://data.42matters.com/stats
S37Tsinghua University: Densing Law of LLMs (Nature Machine Intelligence cover article). Maximum capability density of open-source LLMs roughly doubles every 3.5 months; a model with half the parameters can match the state of the art about every 3.5 months. Dataset: 51 open models, Feb 2023 to Apr 2025. Assumes models trained equally well have the same density; describes the best open models, not typical ones. issue of 20 Nov (year not stated on page); data to Apr 2025older than 1 year https://www.tsinghua.edu.cn/en/info/1245/14611.htm
S38BigGo Finance report on a Counterpoint Research forecast (Kang Kyung-soo, iForum 2026, 22 Oct 2026): new memory output from SK Hynix and Micron from 2027, Samsung and CXMT expansions ramp in 2028; headline cites prices reversing to a 20-30% decline by 2028. Same article: TrendForce expects consumer memory prices to keep rising next year; Q3 2026 DRAM contract prices +13-18% quarter on quarter. 22 Oct 2026 https://finance.biggo.com/news/c3a3fef8-8dfc-4a8b-9acf-c3ae851ea8c8
S40God Is A Geek: RTX 6090 rumoured for H1 2027 (Moore's Law Is Dead, 14 Sep 2026); leakers dispute it and suggest 2028; up to 32 GB GDDR7 in unattributed leaks. NVIDIA has announced nothing. Sep 2026 https://godisageek.com/2026/09/nvidia-rtx-6090-h1-2027-rumor/
S44Canalys, as reported by Vivre and Computer Weekly: AI-capable PCs were 14% of global shipments in Q2 2024 (8.8M units) and 20% in Q3 2024 (13.3M units). Definition: a desktop or notebook with a chipset or block for dedicated AI workloads, such as a neural processing unit (NPU). Aug and Nov 2024older than 1 year https://www.computerweekly.com/news/366615506/A-fifth-of-new-PCs-shipped-this-quarter-are-AI-optimised
S45Tom's Hardware, 'What is an AI PC': Microsoft Copilot+ PC requires an NPU of at least 40 TOPS, 16 GB RAM and 256 GB storage; Intel defines an AI PC as having a CPU, GPU and NPU with AI acceleration. The article does not say whether these PCs can run large language models locally. 15 Jun 2024older than 1 year https://www.tomshardware.com/laptops/what-is-an-ai-pc
S46GitHub topic pages, public repositories matching the topic: 'ollama' 5,794 and 'local-llm' 9,477 (page counts, 25 Oct 2026). Counts include libraries, tools and forks, not only end-user apps; overlap between topics not removed. 25 Oct 2026 https://github.com/topics/local-llm
Key takeaway. Capable hardware is scarce and getting costlier; applications are many, uneven and hard to compare; the capability gap to frontier is small for giant models and larger for models a normal machine can run. All three are measurable each quarter, which is what the series should track.
자유로운 인용과 재사용
이 보고서의 문장과 수치는 모두 인용할 수 있으며, 차트와 표도 기사, 뉴스레터, 강연, 소셜 게시물에서 별도 허락 없이 재사용할 수 있습니다. PDF와 CSV 파일도 마찬가지입니다. 라이선스는 CC BY 4.0입니다. 요청 사항은 다음과 같습니다. 저자와 PromptQuorum를 밝히고, 이 페이지로 링크를 걸고, 내용을 바꾼 부분이 있으면 그 사실을 밝혀 주십시오.
이는 PromptQuorum의 텍스트, 수치, 차트, 표에 해당합니다. 수치의 바탕이 되는 제3자 출처([S…] 참조)는 각 권리자의 약관을 따르므로, 그 출처는 원출처에서 인용해 주십시오. 수치는 보고서마다 달라지므로, 수치를 인용할 때는 출처 표기와 데이터 기준일(2026년 10월 23일)을 함께 적어 주십시오.
<a href="https://www.promptquorum.com/ko/about/state-of-local-ai/2026-q4">로컬 AI 현황, 보고서 01 (Q4 2026) – PromptQuorum</a>
문의, 인용 확인, 수치 정정 요청: hello@promptquorum.com. 이 보고서를 바탕으로 글을 게시하셨다면 링크를 보내 주십시오. 이곳에 소개하겠습니다.
닭과 달걀 심볼, 그리고 누구를 위한 것인지
이 보고서에서 닭은 좋은 로컬 AI 모델을 돌릴 수 있는 컴퓨터이고, 달걀은 쓸 만한 앱입니다. 심볼은 직접 쓰는 기사, 슬라이드, 게시물에 가져다 쓰셔도 됩니다. 라이선스는 CC BY 4.0입니다. 저자와 PromptQuorum를 밝히고, 이 페이지로 링크를 걸고, 내용을 바꾼 부분이 있으면 그 사실을 밝혀 주십시오. "닭이 적다"와 "달걀은 더 적다" 그래픽은 일러스트이지 차트가 아니며, 어떤 데이터도 나타내지 않습니다.
어떤 독자이신가요?
달걀 찾는 사람 · 일반 사용자앱을 내려받아 터미널 없이 쓰고 싶으신 분입니다. 궁금한 것은 찾아갈 만한 달걀이 과연 있느냐입니다.앱과 지불 의향다운로드닭 키우는 사람 · 개발자이미 모델을 돌리고 있고 터미널이 익숙하신 분입니다. 이 보고서는 개발자를 초기 시장으로 봅니다(보고서가 가정으로 표시한 내용입니다).하드웨어 기반 · 로컬 모델과 최첨단 모델다운로드닭장 관리자 · 기업 IT 담당자기기 대수와 예산 단위로 구매하시는 분입니다. 하드웨어 가격, 무역과 공급, 가정을 밝힌 전망이 해당됩니다.하드웨어 기반 · 무역과 공급망 · 전망다운로드달걀 검수자 · 기자 또는 애널리스트수치를 인용하고 출처를 확인하시는 분입니다. 주장에는 증거 표시가 붙고, 수치는 번호가 붙은 출처에서 가져왔습니다.프레스 키트 · 주장 목록 · 출처다운로드
Kuepper, H. (2026). Few Chickens and Even Fewer Eggs. The State of Local AI, Report 01 (Q4 2026). PromptQuorum. https://www.promptquorum.com/about/state-of-local-ai/2026-q4
이 보고서에 대하여
필자는 로컬 AI 앱을 다루는 콘텐츠 사이트이자 디렉터리인 PromptQuorum를 설립했으며, 그곳에 등재된 앱을 테스트했습니다. 증거 레이블과 테스터 본인의 실수를 포함한 각 직접 테스트 기록은 증거 파일에서 링크로 연결됩니다. 테스트 노트에는 대가 지급이나 후원이 기록되어 있지 않습니다.
Hans Kuepper 는 프랑크푸르트에 기반을 둔 PromptQuorum의 창업자로, 30년 넘게 컨설팅과 경영 일선에서 일해 왔습니다.
