marketdeep dive

    ⛵ Dhow Dispatch | What If Intelligence Becomes Free?

    Issue 31 · Focus: AI, China & the Economics of Intelligence

    July 30, 202612 min read
    View original
    ⛵ Dhow Dispatch | What If Intelligence Becomes Free?

    Download the app. Join the community. Back the builders shaping our future.

    Dhow - Private Markets App - App Store

    Download Dhow - the community-powered private markets platform helping Muslim investors discover opportunities, learn, connect, and back Muslim founders building the future.

    Dhow - Private Markets App - App Store

    Download Dhow - the community-powered private markets platform helping Muslim investors discover opportunities, learn, connect, and back Muslim founders building the future.

    Our $10,000 allocation spots are now fully filled, and we expect to begin deploying capital shortly. Our first investment will be in the consumer AI space. The minimum commitment for new investors is now $25,000, with future access expected to become increasingly limited. Reserve allocation here.

    🚀 TL;DR

    For most of the AI boom, frontier intelligence looked scarce.

    Building the best models required enormous clusters, elite research teams and staggering amounts of capital. The companies that controlled those models could charge accordingly.

    China has spent the past eighteen months attacking that assumption.

    DeepSeek, Moonshot AI, Z.ai, Alibaba and others have pushed capable models into the market with downloadable weights, permissive licenses, aggressive pricing and architectures built around squeezing more intelligence out of limited compute.

    The pressure is now spilling into the American market.

    Today, OpenAI cut the price of GPT-5.6 Luna by 80%, from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens. Terra dropped 20% to $2/$12. OpenAI told Reuters that efficiency gains helped make the cuts possible.

    Meanwhile, DeepSeek V4 Flash currently charges just $0.14 per million uncached input tokens and $0.28 per million output tokens.

    The emerging question is bigger than China versus America:

    ❝

    What happens to technology when intelligence becomes abundant infrastructure?

    Our view is that value starts migrating away from the model itself and toward the things that remain difficult to copy: distribution, proprietary data, customers, trusted workflows, physical infrastructure and execution.

    🇨🇳 | DeepSeek was the warning shot

    In January 2025, DeepSeek forced investors to confront an uncomfortable possibility.

    Maybe frontier AI did not require the level of spending everyone had assumed.

    Nvidia fell 17% on January 27, wiping out roughly $593 billion in market value in one day, then a Wall Street record. Investors were reacting to the possibility that increasingly capable AI could be built and operated with dramatically fewer resources.

    One number dominated the headlines: $5.576 million.

    DeepSeek reported that the official training of V3 consumed 2.788 million Nvidia H800 GPU-hours. At an assumed rental rate of $2 per GPU-hour, that works out to $5.576 million.

    That figure is real.

    Its interpretation was often wrong.

    DeepSeek explicitly said it excluded prior research, architecture experiments, data work and other development costs. It represented the estimated compute cost of the reported training run, rather than the full cost of inventing DeepSeek V3.

    The more meaningful result was the engineering.

    DeepSeek used a Mixture-of-Experts architecture, memory-efficient attention, lower-precision training and hardware-aware optimization to extract more capability from available compute. Its later research has continued emphasizing the relationship between model architecture and hardware constraints.

    That approach did not stay confined to DeepSeek.

    It became part of a much broader Chinese model race.

    🧠 I China now has an ecosystem

    DeepSeek V4 arrived in April.

    The Pro version has 1.6 trillion total parameters with 49 billion active, while Flash uses 284 billion total and 13 billion active. Both support a 1-million-token context window, and DeepSeek released downloadable weights. Its official API now charges $0.435/$0.87 per million uncached input/output tokens for Pro and $0.14/$0.28 for Flash.

    Moonshot AI followed with Kimi K3 in July.

    K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion active parameters, native vision and a 1-million-token context window. Moonshot released its full weights on July 27. Its own technical report says K3 still trails the strongest proprietary systems overall, while reaching frontier-level performance across several coding, agentic and reasoning evaluations.

    K3 is also useful because it complicates the cheap-China narrative.

    Moonshot prices its official API at $3 per million uncached input tokens and $15 per million output tokens. This is far above DeepSeek V4 Flash and, following today's OpenAI cuts, far above GPT-5.6 Luna.

    The important trend is therefore broader than price.

    Chinese labs are making increasingly capable intelligence portable.

    Z.ai released GLM-5.2 under the MIT license with a 1-million-token context window. Its Hugging Face repository showed more than one million downloads over the preceding month when checked this week. Download counts are an imperfect adoption metric, but the scale shows meaningful developer interest.

    Alibaba took a similar approach with Qwen3. The flagship Qwen3-235B-A22B was released under Apache 2.0, alongside a family of smaller downloadable models.

    The strongest versions of these systems are still expensive to train and host.

    But once the weights exist, thousands of developers can use them without rebuilding the original research effort.

    That changes the economics of innovation.

    📊 | The Price of Intelligence Is Collapsing

    DeepSeek’s V4 models are pushing hosted inference toward commodity pricing, while OpenAI’s Luna cut shows the pressure is spreading across the market. The cost gap is narrowing fast enough to change how companies think about model choice, routing, and margins.

    💸 | Then America joined the price war 

    This may be the most important development in the entire story.

    When GPT-5.6 launched on July 9, OpenAI priced Luna at $1 input and $6 output per million tokens. Terra was $2.50/$15.

    Three weeks later, those prices have already changed.

    OpenAI cut Luna by 80% today. Terra came down 20%. Sol, the flagship tier, stayed at $5/$30. Reuters reported that OpenAI attributed the cuts partly to efficiency improvements and said companies were increasingly scrutinizing the cost of AI workloads.

    Think about the pace of that.

    A newly released American model lost four-fifths of its API price in twenty-one days.

    At the new rate, 100 million Luna output tokens cost $120.

    DeepSeek V4 Flash would charge $28 for the same token count.

    That does not mean the two systems deliver equivalent output. It shows how far the unit price of machine intelligence has fallen.

    And there is another dynamic here.

    As models become more capable, applications can route different jobs to different systems.

    A hard research task can go to the expensive model.

    Routine classification can run on something cheaper.

    Agents running thousands of background steps can select models dynamically based on cost, latency and capability.

    The application owns the customer.

    The underlying intelligence starts looking increasingly interchangeable.

    🧠 | What happens when intelligence becomes cheap?

    If strong models remain broadly available at falling prices, access to intelligence becomes a weaker competitive advantage.

    Twenty companies can build on DeepSeek.

    Twenty can use Qwen.

    Twenty can call OpenAI.

    Only one may own the customer.

    That shifts the moat.

    A healthcare AI company with years of proprietary clinical data, hospital integrations and established trust owns assets its underlying model provider cannot simply reproduce.

    The same logic applies to financial workflows, industrial systems, logistics networks, legal software and enterprise operations.

    Distribution becomes more valuable because the company controlling the customer can change the model underneath the product.

    Proprietary data gains importance because general intelligence does not automatically possess a company's private history, domain-specific outcomes or internal operating data.

    Workflow ownership becomes sticky once AI is embedded into the systems where work actually happens.

    Physical infrastructure stays scarce. Chips, datacenters, energy systems, factories, warehouses and robots still require enormous capital and time.

    And efficiency may increase demand for that infrastructure.

    When the cost of using a resource falls, usage can rise. That is the logic behind Jevons paradox. Nvidia made a similar argument after DeepSeek's 2025 breakthrough, saying advances in reasoning models would ultimately require substantial inference capacity as adoption expanded.

    A model might use less compute per task while the world generates far more tasks.

    Cheap intelligence could therefore shrink the value of the model layer while expanding the total market built around it.

    📊 | If Intelligence Becomes Abundant, What Stays Scarce?

    As intelligence gets cheaper, durable value shifts toward what remains scarce: distribution, proprietary data, trust, customer relationships, infrastructure, and execution.

    🌍 | The geopolitical opportunity

    There is another layer to this.

    Since October 2022, the United States has repeatedly tightened restrictions on China's access to advanced computing chips and semiconductor-manufacturing technology. The rules expanded in 2023 and 2024 to cover additional advanced chips, equipment, software tools and high-bandwidth memory.

    Those controls made compute harder to access.

    They also made efficiency strategically valuable.

    A June 2026 paper examining U.S. policy shocks found that Chinese developers increased engagement with open-model repositories substantially more than American developers after major export-control events. The authors argue that the restrictions increased the strategic value of open and locally adaptable AI inside China. The study shows an association and strategic response, rather than proving that export controls caused China's open-model boom.

    China is also building more of the hardware stack domestically.

    Huawei's Ascend 950 infrastructure now supports DeepSeek V4. Reuters reported in April that ByteDance, Tencent and Alibaba were seeking Ascend chips after the V4 launch, even though Huawei's hardware still faces manufacturing constraints and trails Nvidia's most advanced systems.

    The geopolitical result reaches far beyond China.

    Governments and companies increasingly want control over where models run, where sensitive data lives and how dependent they are on foreign providers.

    Reuters reported this week that open-weight AI is becoming attractive as a partial hedge against dependence on U.S. proprietary platforms, with India, South Korea and Japan investing in domestic or sovereign AI strategies.

    That opportunity is especially interesting across the Muslim world.

    The Gulf has capital and an explicit interest in sovereign infrastructure.

    Pakistan, Egypt and Indonesia have huge populations and deep technical talent pools without Silicon Valley's historical access to venture capital and compute.

    Malaysia and Indonesia have massive local-language digital markets.

    A founder in Lahore, Jakarta or Cairo increasingly begins with sophisticated intelligence already available.

    Their scarce resources can go toward product, customer acquisition, proprietary data and distribution.

    That narrows one of the historical gaps between founders in capital-rich ecosystems and everyone else.

    🔭 | What we’re watching

    The frontier gap. Open models do not need to beat the best proprietary system at every benchmark. They need to remain close enough for most commercial workloads that the premium becomes difficult to justify. Kimi's own K3 paper says the model still trails the strongest closed systems overall, which makes this an open question rather than a settled outcome.

    Infrastructure demand. Efficiency could reduce compute required per task while cheaper AI simultaneously drives an explosion in total usage.

    Policy. The U.S. is already debating how Chinese open models should be treated. Nvidia, Microsoft, Meta, IBM and other companies backed a July 24 letter opposing broad restrictions on open-weight AI, while Washington considers sanctions and other measures targeting Chinese developers amid accusations of technology theft.

    🧭 | Bottom Line


    The most important AI shift may turn out to be economic.

    Models will keep getting better.

    At the same time, the cost of using them keeps falling.

    DeepSeek showed how far efficiency could be pushed.

    Chinese labs turned open-weight AI into a serious global ecosystem.

    Now OpenAI has cut Luna's API price by 80% only weeks after launch.

    China helped accelerate the race toward cheap intelligence.

    The race is now global.

    If strong model capability continues moving toward commodity pricing, simply owning a powerful model becomes a weaker moat.

    The durable value shifts toward companies that own customers, data, critical workflows, distribution and infrastructure.

    For founders, sophisticated software gets cheaper to build.

    For investors, the question becomes much more interesting:

    When everyone has access to intelligence, who owns everything around it?


    At Dhow, we back builders who chart new waters. Cheap, open AI is collapsing the cost of building and giving founders in places like Lahore, Jakarta, Cairo, Dubai, and beyond access to capabilities that once required massive teams and capital. The edge now comes from distribution, proprietary data, trust, and execution. When the model becomes interchangeable, the company around it becomes the moat. That’s where categories get built and value compounds. Join the movement, share this with a friend (or two).

    Stop typing what you could say in 10 seconds.

    Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

    If you enjoyed this, you’d be doing us a HUGE FAVOR by just clicking the ad above and below - we appreciate it!

    Hiring in a new country? Read this first.

    Every country has different employment laws, payroll requirements, notice periods, and benefits expectations.

    Oyster's Global Hiring Guides help you navigate the details, avoid surprises, and hire with confidence—wherever you're growing next.

    Sources

    Enjoyed this article?

    Subscribe to Dhow Dispatch for weekly insights on values-aligned investing and Muslim entrepreneurship.