Weekly highlights

The Freightos Weekly Update is on hiatus this week – but we’ll be back next week!

In the meantime, here are this week’s changes to freight rates on some of the major lanes.

Ocean rates – Freightos Baltic Index

Asia-US West Coast prices (FBX01 Weekly) decreased 12% to $6,212/FEU.

Asia-US East Coast prices (FBX03 Weekly) decreased 1% to $9,002/FEU.

Asia-N. Europe prices (FBX11 Weekly) decreased 3% to $5,575/FEU.

Asia-Mediterranean prices (FBX13 Weekly) decreased 2% to $6,697/FEU.

Air rates – Freightos Air Index

China – N. America weekly prices decreased 2% to $5.76/kg.

China – N. Europe weekly prices decreased 10% to $3.84/kg.

N. Europe – N. America weekly prices stayed level at $1.93/kg.

Freightos Terminal: Real-time pricing dashboards to benchmark rates and track market trends.

The post Freight rate update for July 29th – July 29, 2026 Update appeared first on Freightos.

For the past several years, the artificial intelligence infrastructure market has followed a simple imperative: build as much computing capacity as possible, as quickly as possible.

Cloud providers ordered graphics processing units in extraordinary volumes. Technology companies committed billions of dollars to new data centers. Utilities received requests for power loads comparable to those of entire cities. Investors rewarded companies positioned anywhere along the AI infrastructure supply chain.

That expansion is not ending. What is changing is the standard by which it will be judged.

The first phase of the boom was defined by scarcity. Companies needed access to advanced chips, networking, cloud capacity, power, land, and technical talent. The strategic risk was failing to build enough capacity while competitors moved ahead.

The next phase will be defined by utilization, economics, and execution.

Investors, customers, and corporate boards will ask harder questions. How much capacity is productive? Which workloads justify premium computing resources? How quickly can new facilities be energized? Where is the revenue? Who bears the risk when technology, demand, and infrastructure timelines do not align?

AI capital spending is becoming a supply-chain and asset-productivity challenge.

The Spending Has Not Stopped

The largest cloud and technology companies continue to spend heavily on data centers, servers, networking, power systems, cooling, and specialized processors.

But scale does not guarantee that every investment will earn an adequate return.

During the first stage of generative AI adoption, access to computing capacity was itself a competitive advantage. Advanced accelerators were scarce, cloud providers rationed access, and companies paid premium prices to train models and launch services.

Under those conditions, the case for more infrastructure appeared almost self-evident. If capacity could be built, demand would likely fill it.

That assumption is now being tested.

AI-related cloud revenue is growing, but the infrastructure required to support it is expanding even faster. This does not prove that spending is excessive. Infrastructure investment often precedes revenue by years.

It does mean that the burden of proof is changing. Management teams must show not merely that AI demand exists, but that expensive assets can be deployed, utilized, and monetized quickly enough to support their financing, depreciation, operating, and energy costs.

Transformative Technology Can Still Produce Bad Investments

The debate is often framed too simply. It is not a choice between believing that AI will transform the economy and believing that some infrastructure will be overbuilt. Both can be true.

AI adoption may continue to expand while some data centers, cloud contracts, and financing structures produce disappointing returns. Capacity may be built in the wrong locations, arrive before sufficient power is available, or become economically outdated faster than expected.

The decisive questions are more specific:

Which AI workloads will generate durable demand?

How much capacity will support training versus inference?

How quickly will computing efficiency improve?

Which providers will retain pricing power?

How long will current hardware remain economically competitive?

Inference Must Sustain the Next Phase

The early infrastructure boom was propelled by training increasingly large foundation models. Training requires enormous accelerator clusters, high-bandwidth networks, large datasets, and sophisticated cooling and power systems.

Inference—the operation of trained models to answer questions, analyze data, control agents, and support applications—could create a broader and more durable market. Its economics, however, are different.

Enterprise users care about latency, reliability, privacy, accuracy, and cost per transaction. They do not need the largest model or most advanced processor for every task.

A supply-chain application classifying documents may require far less computing power than an agent analyzing a global disruption. A forecasting assistant may combine machine learning, retrieval, optimization, and a smaller language model rather than invoke a frontier model for every interaction.

As enterprise architectures mature, companies will become more selective about where they use premium computing.

The next phase will reward providers that match each workload to the appropriate model, processor, memory configuration, network, and service level. The objective will shift from maximizing raw compute to optimizing useful compute.

Utilization Becomes the Critical Metric

AI infrastructure is expensive before it produces a single output.

A functioning cluster requires accelerators, servers, memory, networking equipment, storage, power distribution, backup generation, cooling, buildings, land, security, and connections to telecommunications and electricity networks.

Advanced processors may remain technically useful for years, but their economic attractiveness can decline quickly when new generations deliver better performance per watt or lower inference costs.

An underutilized warehouse can sometimes wait for demand. An underutilized AI cluster may become less competitive while it waits.

Providers need enough capacity to meet bursts of demand and maintain reliability, but idle accelerators still consume capital. High utilization improves economics, while utilization near physical limits reduces flexibility and complicates maintenance.

It will also create demand for better workload orchestration. Computing tasks may be scheduled according to urgency, energy availability, service levels, processor type, and location. Noncritical workloads may shift to periods when electricity is cheaper or the grid is less constrained.

Power Is Becoming the Primary Constraint

The AI industry can manufacture more chips, raise more capital, and build larger models. It cannot instantly create transmission lines, substations, transformers, generating capacity, and grid interconnections.

The immediate problem is regional concentration. Data centers may represent a manageable share of global electricity demand while placing severe pressure on particular local power systems.

Site selection is therefore becoming a strategic sourcing decision.

The best location is no longer simply the one with inexpensive land, favorable taxes, and fiber access. Developers must evaluate electrical capacity, interconnection timelines, transmission constraints, long-term power pricing, water availability, weather exposure, permitting, community opposition, labor availability, and proximity to users and networks.

In some cases, power supply must be developed alongside the data center. Technology companies are examining renewable energy, batteries, natural gas generation, nuclear power, utility agreements, and dedicated generation.

Regions that can provide reliable power, equipment, permitting, and skilled labor may attract the next generation of AI investment. Those that cannot may lose projects regardless of their technology talent or access to capital.

Data Centers Are Becoming Industrial Megaprojects

The scale of proposed AI campuses is moving them beyond the traditional data-center model.

A multi-gigawatt campus resembles a major industrial development requiring coordination among technology providers, utilities, equipment manufacturers, construction firms, financiers, regulators, and communities.

These projects combine large capital requirements, long equipment lead times, interdependent schedules, changing technical specifications, complex permitting, scarce specialized labor, and uncertain demand forecasts.

A delay in one element can strand the others. A facility may be structurally complete but unable to secure electricity. Processors may arrive before cooling systems are ready. Grid equipment may be delayed. New chips may require changes to power density or thermal architecture.

The supply chain must therefore be managed as an integrated program rather than a sequence of independent procurement decisions.

Financing Will Face Greater Scrutiny

The largest cloud companies can finance substantial investment from their balance sheets. But the scale of the buildout is creating more complex arrangements among infrastructure funds, developers, utilities, chip suppliers, sovereign investors, cloud providers, and AI companies.

A developer may build the facility. A utility may finance grid upgrades. A cloud provider may sign a long-term lease. An AI company may commit to capacity it expects future customers to consume.

Customers may renegotiate, delay deployment, encounter financing problems, or find that technological progress changes their capacity requirements. Investors must therefore ask who guarantees the commitments, who funds power upgrades, what happens if energization is delayed, and whether the facility can serve another customer.

These are infrastructure-finance questions, not merely technology questions.

Enterprise Buyers Will Become More Disciplined

Many enterprises have moved beyond experimentation but have not yet achieved broad production deployment. They are focusing more closely on return on investment, governance, workforce readiness, and the practical requirements of scale.

Supply-chain organizations face a demanding business case.

An AI assistant that drafts marketing copy can generate value even when its output is imperfect. An AI agent changing replenishment parameters, selecting a carrier, releasing an order, or responding to a disruption operates in a much less forgiving environment.

Companies must connect AI to trusted data, transactional systems, optimization engines, decision rules, and human approval structures. The strongest use cases will be tied to measurable outcomes such as lower transportation expense, reduced planner workload, faster exception resolution, improved inventory availability, lower expedite costs, reduced downtime, and higher warehouse productivity.

The next phase will favor projects that produce operational results rather than merely demonstrate generative capabilities.

The Opportunity Is Moving Up the Stack

The first infrastructure wave rewarded semiconductor designers, foundries, memory manufacturers, networking suppliers, server vendors, and cloud providers.

The next layer of value will increasingly come from making infrastructure productive.

That includes workload orchestration, model routing, inference optimization, energy management, cloud cost control, AI governance, enterprise integration, and industry-specific applications.

The enterprise does not ultimately want computing capacity. It wants better decisions and improved execution.

A company may use multiple models, cloud providers, private environments, specialized processors, and agent frameworks. The winning platforms will coordinate those resources while maintaining security, visibility, and economic control.

This Is a Transition, Not a Collapse

There are legitimate reasons to be cautious. Capital requirements are rising. Power constraints are real. Some companies will struggle to turn capacity into profitable services. Hardware may depreciate economically faster than expected, and financing structures may weaken if demand assumptions fail.

But increased scrutiny does not mean AI infrastructure demand is disappearing.

The market is moving from indiscriminate expansion toward differentiation.

Projects with reliable power, strong customers, flexible architectures, disciplined financing, and high utilization will remain valuable. Projects built around speculative demand, weak counterparties, unrealistic energization schedules, or inflexible technology assumptions will face greater pressure.

The first phase rewarded access to capital and computing. The next phase will reward execution.

For supply-chain leaders, the lesson is familiar: growth does not eliminate the need for operational discipline. It makes that discipline more important.

The AI infrastructure buildout is continuing, but capacity alone is no longer the objective. The challenge is to convert unprecedented investment in chips, power, and data centers into dependable, economically productive intelligence.

The post AI Infrastructure Is Entering Its Next Phase appeared first on Logistics Viewpoints.

The next major supply-chain dependency may not involve semiconductors, critical minerals, transportation capacity, or industrial components. It may be embedded inside the artificial intelligence models that companies use to build their next generation of operational systems.

Open AI models are becoming foundational components of enterprise technology architectures. Companies can download them, customize them, fine-tune them using proprietary information, and deploy them within private cloud or on-premises environments. This can lower costs, improve control, and reduce dependence on a small number of closed-model providers.

But open does not necessarily mean independent.

As a growing share of the world’s open-model ecosystem consolidates around model families developed outside the United States, companies may be exchanging one form of vendor dependence for another. The result is an emerging strategic question for supply-chain leaders: How much of an enterprise AI architecture should depend on a model ecosystem whose future development, governance, licensing, and geopolitical availability the company does not control?

That question is no longer theoretical.

The Rapid Rise of Qwen

A recent analysis by researchers affiliated with the ATOM Project found a significant shift in the open-model ecosystem. Qwen’s share of newly released fine-tunes and adaptations increased from approximately 1% in January 2024 to 69% by February 2026. Over the same period, Meta’s share declined sharply from its earlier peak.

The numbers do not mean that 69% of all enterprise AI deployments use Qwen. They measure the model families selected by developers when creating new fine-tunes, adapters, and derivative models.

Nevertheless, the trend is strategically important. Fine-tunes and derivative models represent the layer where experimentation becomes application development. They reveal where developers are placing their time, technical knowledge, datasets, integrations, and tooling.

Once a model becomes the default foundation for thousands of downstream applications, it begins to resemble a digital industrial platform. Developers build around its architecture. Software libraries optimize for it. Internal teams develop specialized expertise. Enterprises create evaluation frameworks, deployment pipelines, and governance processes around its behavior.

That produces ecosystem gravity—and ecosystem gravity creates switching costs.

This Is Not an Argument Against Chinese Models

The rise of Chinese open models should not be dismissed as the result of careless adoption or geopolitical naïveté. Many of these models are highly capable, economically attractive, and available under licenses that allow developers to modify and deploy them with considerable flexibility.

Chinese AI laboratories have also demonstrated that strong model performance does not always require the largest possible training budgets. Their progress has increased competition, accelerated open-model development, and placed downward pressure on inference costs.

For enterprises, this is broadly positive.

A manufacturer, retailer, logistics provider, or software company may be able to build useful AI capabilities using models that are less expensive to run and easier to customize than the largest proprietary alternatives. Open models can be deployed closer to operational data, adapted to industry terminology, and integrated into systems where privacy or latency makes a fully hosted service impractical.

The issue is therefore not whether companies should use Chinese-developed models.

The issue is whether companies understand the concentration risk they may be creating when a single model family becomes deeply embedded across their AI stack.

Model Dependence Is Supplier Dependence

Supply-chain organizations already know how to evaluate dependence on a critical supplier. They examine substitution difficulty, geographic concentration, financial stability, production capacity, transportation exposure, regulatory risk, and the time required to qualify an alternative.

A foundational AI model should increasingly be evaluated in the same way.

Consider what happens when a company builds an operational application around a particular model family. The company may create:

Retrieval pipelines optimized for that model

Fine-tuned adapters trained on internal data

Prompt libraries and system instructions

Evaluation benchmarks

Security controls

Model-specific deployment infrastructure

Agent workflows

Integration logic

Employee expertise

Governance and approval processes

The model itself may be freely available, but the surrounding implementation is not free. It represents accumulated investment and organizational learning.

Replacing the model could require revalidating the entire application. Outputs may change. Tool calls may behave differently. Safety controls may need to be redesigned. Fine-tunes may not transfer cleanly. Performance may deteriorate in specialized tasks.

This is exactly what makes a supply source strategically important: not simply the cost of the item, but the cost and operational disruption associated with replacing it.

The Risk Is Broader Than Model Access

The most obvious geopolitical scenario would involve export restrictions, sanctions, licensing changes, or government intervention that affects the availability of a model. But that is only one category of risk.

Enterprises should also consider the broader ecosystem surrounding the model.

Who maintains the underlying architecture? How transparent is the training and post-training process? Where do security updates originate? Which organizations control the primary repositories? How rapidly can vulnerabilities be identified and corrected? What happens when the model’s commercial sponsor changes priorities?

Open weights can reduce dependence on a hosted provider, but they do not eliminate dependence on upstream research, tooling, documentation, and developer communities.

There is also a provenance problem. A company may download a derivative model that has passed through several rounds of fine-tuning by unknown parties. Each stage may alter the model’s behavior, security characteristics, or susceptibility to manipulation.

This does not mean derivative models are inherently unsafe. It means enterprises need stronger software-supply-chain disciplines for AI.

A model should have a traceable lineage. Organizations should know where it originated, who modified it, which datasets were used when that information is available, how it was evaluated, and whether the artifact has been altered since publication.

The same principles behind a software bill of materials will increasingly apply to models, adapters, embeddings, agent tools, and AI-generated code.

AI Sovereignty Is an Architectural Question

AI sovereignty is often discussed at the national level. Governments want domestic access to computing infrastructure, semiconductors, data, talent, and foundational models.

But sovereignty also matters at the enterprise level.

An organization has greater AI sovereignty when it can preserve operational continuity, move between model providers, control its proprietary context, and replace components without rebuilding the entire system.

This does not require every company to train its own large language model. For most businesses, doing so would be economically irrational.

It does require enterprises to avoid architectures in which the model becomes inseparable from the application.

The model should be treated as a replaceable intelligence component rather than the permanent center of the technology stack.

That means separating the model from:

Enterprise data

Business rules

Workflow orchestration

Tool definitions

Security controls

Evaluation datasets

User interfaces

Audit trails

Operational decision rights

This separation is becoming more practical as AI architectures mature.

Retrieval-augmented generation can keep proprietary knowledge outside the model. Graph-enhanced retrieval can provide structured relationships among suppliers, facilities, products, orders, shipments, and disruptions. Model Context Protocol servers can standardize access to data sources and tools. Agent-to-Agent protocols can help specialized agents exchange tasks and information.

Together, these technologies can create an abstraction layer between the foundational model and the operational system.

The enterprise then owns the context, integrations, workflows, and governance—even when it changes the model providing the underlying reasoning capability.

Supply-Chain Leaders Should Demand Model Portability

Technology teams often evaluate models based on benchmark performance, speed, inference cost, and ease of deployment. Those factors matter, but they are incomplete.

Supply-chain leaders should add portability and concentration risk to the evaluation.

Key questions include:

Can the application operate with more than one model family?

A production system should be tested against at least one credible alternative. This does not mean every model will produce identical results. It means the organization understands the effort required to switch.

Is proprietary knowledge stored outside the model?

The more enterprise knowledge is embedded exclusively within a model-specific fine-tune, the harder it may be to migrate.

Are tools and integrations exposed through standardized interfaces?

Model-specific integration code increases lock-in. Standardized APIs, schemas, and protocols make substitution easier.

Does the company maintain independent evaluation datasets?

Enterprises should own the tests used to determine whether a model performs adequately. Vendor benchmarks are not a substitute for operational validation.

Can the organization trace the model’s provenance?

The company should know the model’s source, version, license, modification history, and approval status.

What is the fallback plan?

Critical AI-supported workflows need a defined alternative, which may include another model, a rules-based process, or a human-controlled operating procedure.

How much of the architecture depends on one geopolitical jurisdiction?

This question should include models, cloud infrastructure, chips, development tools, and key software libraries.

Diversification Does Not Mean Fragmentation

Enterprises should not respond by deploying dozens of models without discipline. Excessive variety creates its own costs, including inconsistent outputs, fragmented governance, duplicated infrastructure, and greater cybersecurity exposure.

The objective is not maximum model diversity. It is controlled optionality.

A company may designate one preferred model for a class of tasks while validating one or two alternatives. It may use smaller specialized models for forecasting explanations, transportation exceptions, supplier-risk analysis, or document processing. It may use a larger proprietary model for complex reasoning while keeping an open model available for continuity, privacy-sensitive workloads, or cost control.

This resembles a well-designed sourcing strategy. The organization does not qualify every possible supplier. It identifies where single sourcing is justified, where dual sourcing is necessary, and where standardization creates more value than redundancy.

The United States Faces an Open-Model Policy Dilemma

The enterprise challenge reflects a larger national issue.

American technology policy has strongly supported advanced semiconductor controls and domestic computing infrastructure. But the open-model ecosystem cannot be secured through hardware policy alone.

If American developers increasingly build on foreign model foundations, the United States may retain strength in chips, cloud platforms, and frontier proprietary models while losing influence over the open development layer.

Restricting access to foreign open models would carry significant costs. It could reduce competition, slow innovation, and push developers toward less transparent distribution channels. It could also disadvantage smaller companies that benefit from inexpensive, capable models.

A more productive response would be to strengthen the competitiveness of American open models.

That could include support for open-model research, shared evaluation infrastructure, high-quality public datasets, secure model-distribution systems, and incentives for universities and companies to release commercially usable model families.

Dean Meyer and Konstantine Buhler, whose analysis helped elevate this debate, have also argued for stronger American capabilities in controlled post-training and model adaptation. That direction deserves attention. The strategic contest may not be determined solely by who trains the largest foundational model, but by who creates the most useful ecosystem for adapting models to real operational work.

The Supply Chain Is Moving Inside the Model Stack

Supply-chain risk management has historically focused on physical flows: materials, components, factories, ports, carriers, and inventory.

Enterprise AI adds a new layer.

The organization now depends on model weights, training frameworks, vector databases, orchestration tools, APIs, data pipelines, agent protocols, and cloud infrastructure. These components form a digital supply chain that can be concentrated, disrupted, compromised, or politically constrained.

The rise of Qwen and other Chinese open models should be understood in that context. Their growth is evidence of technical and commercial success. It is also a warning that the open-model ecosystem is consolidating faster than many enterprises realize.

The correct response is neither prohibition nor complacency.

Companies should use the strongest tools available while preserving the ability to change them. They should treat model provenance as a governance issue, portability as an architectural requirement, and concentration risk as a supply-chain concern.

Open models can reduce dependence on proprietary AI providers. But without deliberate architecture and sourcing discipline, they can also create a new strategic dependency beneath the surface of enterprise operations.

The companies that understand this distinction will not merely adopt AI faster. They will build AI systems that remain resilient when the technology market, vendor landscape, or geopolitical environment changes.

The post Open Models, Strategic Dependencies: Why AI Sovereignty Is Becoming a Supply Chain Issue appeared first on Logistics Viewpoints.