ModCon 2026: why one good AI model is no longer enough

Eight hours of ModCon began with a programming language and ended with video whose pictures and sound were generated together. In between came new accelerators, response prices, server failures and electricity shortages. The full conference recording runs for 8 hours and 13 minutes. Breaks occupy almost two hours; the rest is a keynote, panels and short technical talks.

The mainstage program called this a discussion of a unified AI compute layer. In plainer terms, Modular wants teams to move the same model between different chips without rebuilding all of the software around it.

This was Modular’s own conference, so its performance and savings figures are not an independent comparison. The subject itself is revealing, however. Training a good model is no longer enough. It must also serve thousands of people quickly, reliably and at a tolerable cost.

On July 29, three weeks before the conference, Qualcomm completed its acquisition of Modular. Qualcomm CEO Cristiano Amon was therefore describing the combined company on stage, not a possible partnership. Qualcomm has processors and accelerators for devices and data centers; Modular has software intended to make that hardware usable by model developers.

According to Reuters reporting, the deal was worth nearly $4 billion. Qualcomm did not spend that money on its own ChatGPT equivalent. It bought an attempt to solve a less visible problem: a new chip is of little use to most teams when familiar models and tools are difficult to run on it.

Modular divides its answer into three parts. Mojo is used to write fast operations for CPUs and accelerators. MAX loads and serves models. Modular Cloud sells the same stack as a service. Together, they are meant to provide one route to hardware from NVIDIA, AMD, Google, Amazon and Qualcomm.

NVIDIA, AMD, AWS Trainium, Google TPU and Qualcomm accelerators appear together in Modular’s announcement roundup. The stage presentations showed a clearer distinction. NVIDIA and AMD were already serving live workloads, while Trainium, TPU and Qualcomm appeared as newer ports and directions for further work.

“The same code runs everywhere” therefore needs a qualification. Code can run once in a demonstration, pass a test suite or operate for months under unpredictable demand. Those are three different levels of readiness. The last requires uncommon-operation optimization, failure monitoring and cost measurement on real requests.

At ModCon, the Mojo compiler and tools were opened under Apache 2.0 with LLVM exceptions. The Modular repository also shows that MAX is being opened in parts and that community participation in compiler development is still taking shape. An open language is meaningful, but it does not automatically make the cloud and the entire platform independent of their owner.

The clearest technical explanation came in the Modular Cloud discussion. A model first reads the whole request and establishes its context. That phase is especially compute-hungry. It then writes the answer one piece at a time. Here, memory and repeated access to the stored context are more likely to constrain speed.

When one accelerator pool handles both jobs, some machines inevitably wait. Modular proposes splitting the phases: one pool reads requests, another continues responses, and the context moves between them. Each pool can then grow or shrink independently.

This is where portability becomes more complicated than compilation. The platform must move a system of many machines, not one fast operation, while preserving queues, memory, networking and failure recovery. A new chip can calculate brilliantly and still lose if data takes too long to reach it and return.

Artificial Analysis and SemiAnalysis argued for paying less attention to the price of one token, a small piece of text. In real work, a model does more than write text. It waits for searches, opens pages, runs tools and sometimes repeats a failed step. Generation itself may be the cheapest part of the task.

A more useful measure is the cost of finished work: a resolved support request, an accepted code change or a completed report. That amount includes model calls, external tools, waiting time, failures and manual corrections. A more expensive model can cost less overall when it starts over less often.

Investors on the adjacent panel reached a similar conclusion from the capital side. Money is moving beyond training into memory, networking, electricity and specialized accelerators. Open weights provide more deployment choices, but server and operating costs do not disappear.

Google DeepMind, MiniMax, Reflection AI and Poolside discussed how open weights can move models more quickly to Apple Silicon, server accelerators and new software environments. When the original laboratory has not supported a device, a community may be able to add it.

Access to weights is not a finished service. A strong foundation model is expensive to train; after release, it still has to be compressed, accelerated, updated and tested on every new kind of hardware. Speakers compared training with an experiment factory in which the data, code version and origin of every result must remain traceable.

Failures received separate attention. A server can go offline, an agent can choose the wrong action and a long task can stop halfway through. An operating system must notice, preserve completed work and decide whether it can continue automatically or needs a person.

In the Oracle Cloud and TensorWave discussion, available electrical power mattered more than the number of chips. Speakers said a site in parts of Asia could connect months earlier than one in congested regions of North America. That was their estimate, not a general timetable for every project.

Construction is increasingly financed in advance: a customer signs a multi-year compute contract and the supplier uses that commitment to prepare the facility. While demand exceeds supply, even older accelerators retain value. The newest chip cannot help when no rack, cooling system or power connection is available for it.

Hardware selection therefore does not begin with the record specification. Teams first have to discover what can actually be connected, how machines will exchange data and who will adapt the model. Software portability matters precisely because the ideal hardware may not exist in the required place at the required time.

During the final half hour, MiniMax demonstrated video generation from text, images, existing clips and audio. The system could use first and last frames, transfer motion and style, replace objects, preserve a face and synchronize lips with a voice.

This kind of video is substantially heavier than an ordinary text response. It consumes more memory and compute, while image and sound have to agree in time. The closing demonstration was therefore a useful example of the workload for which Modular is building its common software layer.

The stage naturally showed selected clips. They do not establish long-scene failure rates, generation time or the cost of an accepted result. Open weights allow an independent test, but only with suitable hardware and additional integration work.

A conference like this makes it tempting to start comparing chips. The more useful move is the opposite: take one familiar production task and run it on two hardware types that are actually accessible. Completion time, total cost, failures and manual corrections can then replace stage promises. A Russian project must add payment eligibility, data location, equipment procurement and support — questions ModCon did not answer.

The test can begin with an open model in an accessible cloud, on a rented server or on customer-owned hardware. Detailed logs, spending limits and restricted permissions should be in place first. Then the team should try moving the model and data to another supplier. If everything around them has to be rewritten again, the promised freedom still exists only on the slide.

The phrase “MAX licensing” hides an important difference from Mojo. The new MAX license removes restrictions on device choice, but it does not make all of MAX open-source software. Developers may study and modify the code, but may distribute a modified version only as object code. An application may include only the MAX components that Modular expressly designates as redistributable. A company selling a hosted or managed model-training or inference service to other companies must also obtain approval for and display the Powered by Modular mark. MAX components already released under Apache 2.0 remain under that open-source license.

That is why the author of an independent Forbes account of ModCon is watching whether this license and the planned industry alliance give partners real independence, not merely abstract “terms.” Two other questions remain: independent validation of Modular’s performance numbers and whether the claimed porting economics repeat on the next chip. That is Modular’s central risk. Qualcomm paid nearly $4 billion for a promise to reduce dependence on one hardware vendor. Modular must now prove that moving to another chip is genuinely easier than moving away from Modular itself.