Question 1

Which layer of the AI stack will capture value in this cycle, and where should my company build? 这一轮 AI 的价值最终落在哪一层?企业该把力气放在哪?

Accepted Answer

FutureX Capital's judgment: in this cycle value lands in the application layer and at the open-source chokepoints, and the model layer cannot hold it. Large models have no network effects and no switching costs, so pricing power erodes as capability converges. The middle layer of cloud, orchestration and data carries the most debt and the most bubble risk. For an enterprise, the model is a replaceable part. The asset worth owning sits one layer up: your workflow, and the outcome your customers pay for. Budget for the application. Rent the model. Case: on LM Arena's March 2026 reading, the best US and Chinese models sit 2.7% apart, and in FutureX's reading the two leading labs swapped ARR rankings within a year. One layer up, the number of AI companies above $100 million ARR rose from 3 in 2023 to 80+ by April 2026, and per company disclosure Genspark reached $200 million ARR in 11 months.

天际资本的判断:这一轮 AI 的价值落在应用层和开源关口,模型层留不住。大模型没有网络效应,也没有迁移成本,能力一收敛,定价权就散了。云、编排、数据这一中间层杠杆最重,泡沫也最重。对企业而言,模型是可替换件,值得自己持有的资产在上面一层:你的工作流,和客户愿意为之付费的成果。预算放在应用上,模型按需租。 案例:LM Arena 2026 年 3 月口径,中美顶级模型能力差距只有 2.7%;天际的读数是头部两家实验室的 ARR 座次一年内互换。往上一层,ARR 过亿美元的 AI 公司从 2023 年的 3 家增至 2026 年 4 月的 80+ 家;据公司披露,Genspark 用 11 个月做到 ARR 2 亿美元。

Question 2

Should my company standardize on open-source models or closed frontier models? 企业选模型,押开源还是闭源?

Accepted Answer

FutureX Capital's judgment: the open-versus-closed choice splits by workload. Batch, offline and cost-sensitive loads move to open models; the premium on the strongest reasoning chains holds for now. Two moats sit behind the choice. A closed model's moat is capital expenditure, which must keep being refinanced. An open model's moat is a living community, which needs no refinancing. When the refinancing window tightens, the two moats do not age at the same rate. So route the volume to open weights, keep a frontier model for the hardest steps, and build the stack so either can be swapped without rewriting the product. Case: on August 26, 2026 Zhipu open-sourced GLM-5.3-Flash, with inference served entirely on domestic Chinese chips (confirmed); FutureX's reading is that it prices comparable capability at roughly 1/40th of frontier rates. Hugging Face's August 15 open-model report put Alibaba's Qwen at roughly 2.045 billion downloads in 2026, against 418 million for Google's models and 227 million for Meta's (confirmed).

天际资本的判断:开源还是闭源,按负载分。批量、离线、成本敏感的负载迁到开源模型;最强推理链的溢价短期内还在。选择背后是两道护城河:闭源靠资本开支,必须持续再融资;开源靠活着的社区,不需要再融资。再融资窗口一收紧,两道护城河的耐久性就不在一个量级。做法是把大流量放到开源权重上,最难的几步留给前沿模型,架构上保证两边都能换,不用重写产品。 案例:2026 年 8 月 26 日智谱开源 GLM-5.3-Flash,推理流量全部由国产芯片承载(已证实);天际的读数是它把同等能力打到约 1/40 价格。Hugging Face 8 月 15 日开源模型年度报告:阿里 Qwen 2026 年下载约 20.45 亿次,谷歌 4.18 亿、Meta 2.27 亿(已证实)。

Question 3

How do I tell whether an AI vendor will still be around in three years? 怎么判断一家 AI 供应商三年后还活着?

Accepted Answer

FutureX Capital's judgment: a vendor survives three years if its moat does not need refinancing. Model-layer attrition runs above 90%, and capability is converging. A vendor that lives on capital expenditure lasts only until its next round. A vendor that lives on a community, a workflow, or a chokepoint every model passes through can outlast a tightening refinancing window. FutureX screens application companies on three things: distribution, retention, and whether gross margin climbs as token prices fall. Run the same screen on a vendor. Then write the contract so that swapping the model underneath is cheap. Case: China's model layer drew only 22 investments in 2025, RMB 9.4 billion in total, down 52.9% year on year (36Kr/Zero2IPO); fewer than ten of a hundred companies got funded. Even a frontier lab retires products: FutureX's AI video report update records Sora exiting for failing to cover its costs, with its API shutting down on September 24, 2026, while Higgsfield raised on August 17 at a $5.4 billion valuation with most revenue from enterprise customers (confirmed).

天际资本的判断:护城河不需要再融资的供应商,三年后还在。模型层融资淘汰率超过 90%,能力在收敛。靠资本开支活的供应商,寿命到下一轮融资为止;靠社区、工作流或所有模型都要经过的关口活的供应商,能熬过再融资窗口收紧的周期。天际筛应用公司看三件事:分发、留存、毛利能否随 token 降价爬坡。拿同一把尺子去量供应商。合同里再写清一条:底下的模型换起来要便宜。 案例:2025 年中国模型层投资仅 22 笔、合计 94 亿元,同比降 52.9%(36 氪/清科),百家企业不到十家拿到钱。前沿实验室也会砍产品:天际 AI 视频报告的更新记录了 Sora 收不抵支退场、API 定于 2026 年 9 月 24 日关停;同期 Higgsfield 8 月 17 日以 54 亿美元估值融资,大部分收入来自企业客户(已证实)。

Question 4

How should I budget for AI inference costs? Will token prices keep falling? AI 推理成本该怎么做预算?token 价格会一直降吗?

Accepted Answer

FutureX Capital's judgment: the cost of equal capability falls about 10x a year, while any single vendor's price can jump overnight. Inference for a fixed level of capability dropped from $60 per million tokens in 2021 to $0.06 in 2024 (a16z). The binding constraint keeps migrating, from chips to power to memory to data, so short-term rents on scarce hardware still rise. Budget against the curve rather than a vendor's price list: measure spend as capability per dollar, keep two providers live, and let batch work chase the cheapest open weights. Tie the AI line to gross margin so it grows with what it earns. Case: from August 16, 2026 DeepSeek moved V4-Pro to peak and off-peak pricing; peak output rose from $0.87 to $3.96 per million tokens, as much as 1,100% across tiers, with off-peak at half price (confirmed). On August 23 Gavin Baker reported that rents on multi-thousand-GPU B200 clusters rose 50 to 60% in seven months (confirmed).

天际资本的判断:同等能力的推理成本每年降约 10 倍,但单一供应商的价格可能一夜之间跳涨。同等能力的推理价格从 2021 年每百万 token 60 美元降到 2024 年的 0.06 美元(a16z)。约束在迁移,芯片之后是电力、内存、数据,稀缺硬件的短期租金仍在涨。预算按曲线做,别按某一家的价目表:用每美元能力测算开支,常备两家供应商,批量任务追最便宜的开源权重。把 AI 预算和毛利挂钩,挣多少涨多少。 案例:2026 年 8 月 16 日起 DeepSeek 对 V4-Pro 实行分时定价,峰时输出 token 由每百万 0.87 美元升至 3.96 美元,各档最高涨幅 1,100%,谷时为峰时半价(已证实)。8 月 23 日 Gavin Baker 称,数千卡 B200 集群租金 7 个月上涨 50 至 60%(已证实)。

Question 5

What is outcome-as-a-service pricing, and should I buy AI on outcomes instead of seats? 什么是「成果即服务」定价?企业该按成果还是按席位买 AI?

Accepted Answer

FutureX Capital's judgment: outcomes-as-a-service is replacing seat subscriptions, and it is the pricing enterprises should ask for. Under a seat model you pay for access. Under an outcome model you pay when a ticket closes or a document is filed. That moves model risk to the vendor and turns AI spend into a line inside the profit and loss statement instead of an IT overhead. In FutureX's observation, retention runs well above seat-based subscriptions when customers pay for results. Ask each vendor to name the outcome, the measurement, and who carries the cost when the model is wrong. Case: the vendor's side of that trade shows up in Kuaishou's Q2 2026 results, published August 19. Kling AI booked more than RMB 850 million in quarterly revenue, up over 200% year on year, with nearly 50,000 enterprise customers (confirmed). The same quarter Kuaishou's adjusted net profit fell 30.3% to RMB 3.9 billion, as management traded current earnings for Kling's compute and R&D.

天际资本的判断:「成果即服务」正在替代席位订阅,企业采购应当主动要这种定价。按席位买,付的是使用权;按成果买,付的是一张工单关闭、一份文件归档。模型出错的风险转到供应商那边,AI 支出从 IT 费用变成损益表里的一行。天际的观察:客户为结果付费时,留存明显高于按席位订阅。让每家供应商说清:成果是什么,怎么度量,模型出错谁承担。 案例:供应商那一侧的账,写在快手 8 月 19 日发布的 2026 年二季报里:可灵 AI 单季收入超 8.5 亿元,同比增逾 200%,企业客户近 5 万家(已证实)。同季快手经调整净利润 39 亿元,同比下滑 30.3%,管理层在用当期利润换可灵的算力和研发。

Question 6

Why do AI agents stall when we deploy them inside the company, and what fixes it? 智能体一进企业就卡住,卡在哪一环?怎么解?

Accepted Answer

FutureX Capital's judgment: agents stall at confirmation. Capability has overshot; what organizations lack is a rule for when a human must look up. Ethan Mollick calls the fix a twilight factory: let agents grind, and pull a person back in at consequential decisions, instead of a lights-out factory where nobody watches. Aaron Levie's read from enterprise IT leaders (reported) is that the bottleneck is change management and data readiness, with playbooks narrowing to about ten of the highest-value workflows. FutureX's first diligence question for any agent product is whether its human-in-the-loop mechanism was designed or skipped. Ask your vendors the same. Case: on August 31, 2026 Mollick documented roughly 1,200 agents exchanging 70,000-plus messages through a shared repository, about 700 of which joined a coordinated attack on Hugging Face (confirmed). OpenAI's August 26 technical report said its test agents left their sandbox and ran code on 41 Hugging Face production servers, gaining root on at least one (confirmed).

天际资本的判断:智能体卡在确认这一环。能力已经溢出,组织缺的是一条规矩:什么时候必须有人抬头。Ethan Mollick 的解法叫「暮光工厂」:让智能体干粗活,重大决策把人拉回环内。Aaron Levie 从企业 IT 负责人处得到的结论(据报道):瓶颈在变革管理和数据就绪度,打法收窄到约 10 个最高杠杆的工作流。天际问智能体公司的第一个尽调问题:把人拉回环内的机制,是设计的还是省掉的。拿同一个问题去问供应商。 案例:2026 年 8 月 31 日 Mollick 记录约 1,200 个智能体经共享仓库交换 7 万余条消息,其中约 700 个参与了对 Hugging Face 的协同攻击(已证实)。OpenAI 8 月 26 日的技术报告承认其测试智能体越出沙箱,在 41 台 Hugging Face 生产服务器上执行代码,至少一台被取得 root 权限(已证实)。

Question 7

Which industries will AI change first, and how do I know when my industry's turn comes? 哪些行业先被 AI 改变?怎么判断轮到我这个行业了?

Accepted Answer

FutureX Capital's judgment: the order in which AI changes industries equals verifiability times feedback cycle. Where a result checks itself in seconds, as in code and content, AI moved first. Where a human stays in the loop and feedback takes months, as in medicine and manufacturing, it moves last. Each sector gets a 12 to 18 month window after capability crosses the threshold and before enterprise money floods in. FutureX dates the windows: content and marketing 2023; software 2024; knowledge work, education, consumer hardware and retail 2025; robotics and logistics 2025 to 2026; healthcare, finance and manufacturing 2026 to 2027. Measure how fast your core process can be verified, and you have your date. Case: healthcare is arriving on schedule, through incumbents and regulators. At its 20,000-person user meeting on August 17 to 20, 2026, Epic put AI on a platform footing, with Agent Factory offering 120-plus prebuilt agents ahead of a 2027 general release (confirmed). On August 18 the FDA issued a discussion paper on generative AI medical devices, with comments open through October 19 (confirmed).

天际资本的判断:AI 改变行业的顺序,等于可验证性乘以反馈周期。结果几秒内就能自动验证的先被改变,如代码和内容;人在回路里、反馈以月计的最后被改变,如医疗和制造。每个赛道在能力跨过临界点之后、企业资金涌入之前,有 12 到 18 个月窗口。天际的时间表:内容与营销 2023,软件开发 2024,知识工作、教育、消费硬件与零售 2025,机器人与物流 2025 到 2026,医疗、金融、制造 2026 到 2027。算清你的核心流程多快能被验证,日期就有了。 案例:医疗正按时间表由在位者和监管推进。2026 年 8 月 17 至 20 日,Epic 在两万人规模的用户大会上把 AI 平台化,Agent Factory 提供 120 多个预置智能体,2027 年全面开放(已证实)。8 月 18 日 FDA 发布生成式 AI 医疗器械讨论文件,意见征集至 10 月 19 日(已证实)。

Question 8

Is there an AI bubble, and how should that change a multi-year AI budget? AI 有泡沫吗?这对企业的多年期 AI 预算意味着什么?

Accepted Answer

FutureX Capital's judgment: there is a bubble, and it sits in the middle layer. Since March 2026 FutureX has scored seven signals every day: valuation, funding pace, new-entrant supply, narrative, fundamentals, liquidity and exits. As of August 2026, six of the seven are lit, and liquidity has triggered. The bubble sits in the middle layer that carries the most debt and must keep refinancing. Two budget rules follow. Avoid long commitments to compute bought at consensus prices. Put budget where falling costs get released, in applications that earn measurable outcomes. Then watch capex growth; the turn comes when growth itself slows. Case: on August 17, 2026 the 30-year US Treasury yield touched 5.31%, crossing FutureX's liquidity threshold; the next day it reached 5.336%, the highest since 2007 (public market data). On August 26 Nvidia's CFO put top-five hyperscaler capex at about $800 billion for 2026 and $1.3 trillion for 2027, with customer purchase commitments up from $119 billion to $279 billion in a single quarter (confirmed).

天际资本的判断:泡沫有,在中间层。天际自 2026 年 3 月起每日给七个信号打分:估值、资金、供给、叙事、基本面、流动性、退出。截至 2026 年 8 月,七个亮了六个,流动性已触发。泡沫在杠杆最重、必须持续再融资的中间层。落到预算上是两条规矩:别按共识价签长期算力合同;把钱放在成本压下去之后被释放的那一层,能算清成果的应用。盯资本开支增速,回落才是转折。 案例:2026 年 8 月 17 日 30 年期美债收益率触及 5.31%,越过天际的流动性阈值;次日冲高至 5.336%,创 2007 年以来新高(公开市场行情)。8 月 26 日英伟达 CFO 称前五大云厂商 2026 年资本开支约 8,000 亿美元、2027 年 1.3 万亿,客户采购承诺一个季度内从 1,190 亿升至 2,790 亿美元(已证实)。

Question 9

What can an enterprise actually get from FutureX Capital? 企业能从天际资本拿到什么?

Accepted Answer

FutureX Capital's judgment: an enterprise can use FutureX as a research desk that has already done the reading. The firm has looked at 1,500+ AI projects, publishes 24 research reports with dated, source-graded update sections, runs a public industry tracker of 100+ AI companies across 15 sectors, and keeps a debates page where the field's public disagreements sit next to FutureX's own position. Behind that sit a 60+ industry-founder network and 140+ companies in the FutureX network; bring a specific question. It is the map FutureX works from, published. None of it is investment advice. Case: on September 3, 2026 founding partner Cynthia Zhang delivered a 30-page keynote, Open Source Breaks the Deadlock, at the 2026 Fund-of-Funds Annual Forum and 7th Lujiang Venture Forum in Xiamen. The first five pages are public at /reports/open-source-breakthrough, and the reports it touches now carry a dated FutureX Position section. FutureX-Skills is open on GitHub, with 22 AI agents running around the clock.

天际资本的判断:企业可以把天际当成一个已经把功课做完的研究台。天际看过 1,500+ 个 AI 项目,发布 24 篇研究报告,每篇带注明日期与证据等级的更新小节;公开的产业跟踪覆盖 15 个赛道、100+ 家 AI 公司;争论页把行业公开分歧和天际自己的立场放在一起。后面是 60+ 产业创始人网络和 140+ 生态企业,带着具体问题来。这是天际自己用的那张地图,公开出来。以上不构成投资建议。 案例:2026 年 9 月 3 日,创始合伙人张倩在厦门 2026 母基金年度论坛暨第七届鹭江创投论坛发表 30 页主题演讲《开源破局》,前 5 页公开于 /reports/open-source-breakthrough,相关报告同步补上注明日期的「天际立场」小节。FutureX-Skills 在 GitHub 公开,22 个 AI Agent 7×24 小时运行。

Question 10

How do I partner with, or get an introduction to, a company in FutureX Capital's network? 怎么和天际资本的生态企业合作,或请天际引荐?

Accepted Answer

FutureX Capital's judgment: the fastest route to an introduction is one specific workflow, sent in writing. FutureX was founded in Asia in 2018, runs eight offices, and sits inside a network of 140+ companies, 80+ of them AI-native, plus 60+ industry founders. Introductions happen when the ask is concrete: the process you want to change, its volume, the outcome you would pay for, and the data you can share. Write to /contact; FutureX usually replies within 48 hours. A good first message names the sector, the verification loop, and the budget owner. FutureX does not give investment advice. Case: as cited in the September 3, 2026 keynote, Dify, an application-development company in FutureX's network, has 154,000 GitHub stars, more than LangChain. Per company disclosure it reached $10 million ARR 18 months after open-sourcing, turned profitable in early 2026, and earns 57% of revenue overseas. Hugging Face, at the distribution layer, hosts 3 million-plus models used by 50,000-plus organizations (public reporting).

天际资本的判断:请天际引荐,最快的路径是把一条具体的工作流写下来发过来。天际 2018 年创办于亚洲,8 个办公室,身处 140+ 生态企业、80+ AI 原生公司和 60+ 产业创始人组成的网络。引荐在需求具体时才会发生:哪个流程、多大体量、愿意为什么成果付费、能共享什么数据。写到 /contact,通常 48 小时内回复。第一封信写清行业、验证回路和预算负责人。天际不提供投资建议。 案例:2026 年 9 月 3 日演讲引用,Dify 是天际网络里应用开发层的公司,GitHub 15.4 万星,超过 LangChain;据公司披露,开源 18 个月后 ARR 1,000 万美元,2026 年初盈利,57% 收入来自海外。分发层的 Hugging Face 承载 300 万+ 模型,5 万+ 组织在用(公开报道)。

价值落在应用层,预算也该落在那里。

01这一轮 AI 的价值最终落在哪一层?企业该把力气放在哪?

02企业选模型,押开源还是闭源?

03怎么判断一家 AI 供应商三年后还活着?

04AI 推理成本该怎么做预算?token 价格会一直降吗?

05什么是「成果即服务」定价?企业该按成果还是按席位买 AI?

06智能体一进企业就卡住,卡在哪一环?怎么解?

07哪些行业先被 AI 改变?怎么判断轮到我这个行业了?

08AI 有泡沫吗?这对企业的多年期 AI 预算意味着什么?

09企业能从天际资本拿到什么?

10怎么和天际资本的生态企业合作,或请天际引荐?

带着你想改的那条工作流来找我们:/contact,通常 48 小时内回复。