Amazon CEO Jassy on $220 Billion in Capital Expenditure: Compute Is Still in Short Supply, Inference Costs Need to Fall by Another Order of Magnitude

[[**Amazon's 2026 capital expenditure is projected to reach $220 billion; compute capacity remains in short supply, with 2026 capacity sold out, 2027 largely booked, and a large share of 2028 quotas locked in; high inference costs are slowing enterprise AI adoption.**

**Key Points**

**1. AI demand is dumbbell-shaped**

Top-tier labs and super-apps are consuming massive compute, while lightweight scenarios such as customer service are being deployed; the vast number of mid-tier production systems have not deeply integrated AI because inference costs are too high, and costs need to fall by another order of magnitude.

**2. In-house chips are lowering inference costs**

Among AWS's top 1,000 customers, 99% use Graviton, which offers about 40% better price-performance than x86; Trainium has a 30%-40% advantage over comparable hardware.

**3. Partnering with both Anthropic and OpenAI**

Anthropic has long collaborated through Bedrock and Trainium; OpenAI models are entering Bedrock, and OpenAI has committed to purchasing several gigawatts of Trainium compute.

**4. AI is compressing teams and forcing rewrites**

This article is from the WeChat public account: AGI 2050, author: AGI 2050, original title: "Amazon CEO In-Depth Interview: Spending $220 Billion a Year, Why Is Compute Still Not Enough?"

The global tech industry is undergoing an unprecedented infrastructure buildout. While the capital markets are still debating whether compute investment is overheating, Amazon, the parent company of the world's largest cloud computing platform AWS, has given a clear guidance: capital expenditure for 2026 alone is expected to reach $220 billion.

This scale already exceeds the asset size of most multinational giants. Yet in the eyes of Amazon's current CEO, Andy Jassy, even with such massive investment, compute supply still cannot keep up with demand—capacity for 2026 and 2027 is essentially sold out, and even a large portion of 2028 quotas has already been locked in.

Jassy is the actual architect of AWS and has been at the helm of Amazon for five years since succeeding Bezos. In the latest conversation hosted by Ben Horowitz, co-founder of the well-known venture capital firm a16z, Jassy responded directly to the various controversies surrounding AI. From accusations about data center energy consumption to the persistently high inference costs for enterprises; from how in-house chips complement Nvidia's ecosystem to the business balance of investing in both Anthropic and OpenAI; and to how small teams inside large enterprises can disrupt traditional development models—the conversation is extremely information-dense and objectively presents how a leading cloud provider envisions the endgame of the AI industry.

This episode comes from The a16z Show podcast, aired on October 23, 2026; the original video is 61 minutes long.

Below are selected excerpts from the conversation, about 9,000 words in total, with a reading time of about 5 minutes.

Highlights from This Episode

●

Even with the $220 billion in capital expenditure we plan to invest this year, we still cannot meet all customer demand. Compute capacity for 2026 is already sold out, 2027 is largely booked, and a large portion of 2028 quotas has already been locked in.

●

There are many misunderstandings in outside criticism of data center energy consumption. The water used by golf courses in two days exceeds the total for all our data centers worldwide for an entire year; the total water consumption of golf courses is 200 times that of data centers.

●

Frontier AI labs are willing to pay extremely high premiums to secure power and chips, but we will never squeeze out traditional core customers who have run stably on AWS for twenty years because of that. By helping customers succeed over the long term, our business will naturally follow.

●

Current enterprise AI demand is typically "dumbbell-shaped": on one end are model labs consuming massive compute, and on the other are simple customer-service cost-reduction applications; the vast majority of truly core enterprise production systems have not yet deeply adopted AI. The core bottleneck hindering deployment is the high inference cost, which must fall by an order of magnitude before a true explosion arrives.

●

When we rebuilt Bedrock's core inference engine, it took only 6 engineers and 76 days to complete work that previously would have taken 40 people a full year to deliver. In the future, the number of people in large enterprises engaged in traditional routine roles will certainly continue to decline.

●

Rather than focusing only on more bizarre scenarios like "AI destroying humanity," it is better to focus on the most pressing threat right now—cybersecurity. Large models can find many vulnerabilities that humans had not previously discovered; if an enterprise's codebase cannot be automatically fixed and deployed by agents within minutes, it simply cannot withstand future attacks.

●

Q: Jassy, it has been five years since you took over as Amazon CEO. Amazon's capital expenditure this year is expected to reach $220 billion, and AWS has returned to a high-growth track. Looking back at when you took over in 2021, how has your strategic thinking changed?

Jassy: 2021 already feels very distant now. If we compare the operating data over these five years, our overall revenue grew by about 80%, AWS's business scale expanded to nearly three times, and the advertising business is two and a half times what it was then.

These five years have not been easy. When I first took over, it was the tail end of the pandemic, and the e-commerce business ran head-on into sharply rising fulfillment costs. Within 18 months after the pandemic broke out, we doubled the warehousing and logistics network that had taken 25 years to build; then, within another 20 months, we built a delivery network equivalent in scale to UPS.

Expanding at that extreme speed inevitably created a large number of unoptimized links, and operating costs at one point far exceeded expectations. At the time, many outsiders believed that taking over at that point was "catching the worst possible timing."

But in hindsight, it was actually a rare opportunity. That pressure forced us to break the underlying assumptions we had long adhered to. In the past, we operated a flat, nationwide logistics network in the United States; to compress costs and improve speed, we resolved to split it into eight independent regional networks. This brought inventory closer to users and greatly shortened transport distances, which not only lowered fulfillment costs but also substantially increased delivery speed, in turn driving a new round of rapid growth in the retail business.

Then generative AI exploded. Although Amazon had invested in machine learning for many years, the large-model wave since 2023 has been completely different. Previously we expected AWS to grow into a business with hundreds of billions of dollars in annual revenue within a decade; now, as all digital experiences need to be reshaped by model inference, the growth space has been opened up further. Last quarter, on the basis of $169 billion in annualized revenue, AWS still achieved 36.7% year-over-year growth. Combined with low-Earth-orbit satellite networks, autonomous driving, and robotics, the customer scenarios we can serve are far richer than before.

Q: Doubling the logistics network in an extremely short period back then inevitably came with chaos and inefficiency. Amazon has always emphasized "frugality" and operational excellence. As CEO, how did you maintain the company's organizational culture and execution standards while pushing aggressive expansion?

Jassy: Amazon has 16 leadership principles, and outsiders often assume they are fully consistent, but in reality many of them are in natural tension with one another.

The core choice facing management at the time was this: if we slowed the pace, the plan would be more mature and costs more controllable; but during the special period when physical commerce was largely shut down, consumers and third-party sellers across the United States relied heavily on Amazon. We had to choose between "meticulous planning" and "meeting customers' urgent needs." We ultimately chose "bias for action," prioritizing our ability to fulfill orders for customers and merchants.

In rapid execution, design deviations are inevitable. But Amazon has an internal culture of "facing the truth": we are used to putting real data and business problems directly on the table, without glossing over them, and discussing issues bluntly. Once the main body of the infrastructure was in place, we immediately put the previously deprioritized network architecture reorganization on the agenda, and through systematic architectural adjustments brought costs and speed back into the optimal range.

Data center controversy and the misread energy ledger

Q: Now let's look at today's AI infrastructure. As computing demand surges, data centers have triggered widespread social and policy controversy around the world, especially in the United States. Capital markets are beginning to question whether global funds are sufficient to support this round of infrastructure, and ordinary people are also worried about high energy consumption. Is there any basis for these voices of opposition to data centers?

Jassy: First, we need to establish a basic consensus: the normal operation of modern society—whether financial transactions, medical systems, online retail, or government affairs and travel—is entirely built on digital infrastructure, and data centers are the foundation of these systems. If you cannot live without smartphones and digital services on one hand, yet wholly resist data centers on the other, that is logically untenable.

More importantly, AI is the most significant technological change of this generation, and the scale of computing power it requires represents an infrastructure undertaking unlike anything since the United States built the interstate highway network in the 1950s. Whether from the perspective of industrial upgrading or international technological competition, failing to develop data centers would create enormous strategic vulnerability.

Of course, as infrastructure, the industry does need to keep improving on environmental protection and operational efficiency. But much of current public opinion contains misreadings that depart from the facts:

First, water consumption. Many people accuse data centers of being "water guzzlers," but the fact is that golf courses across the United States, or California's almond farming industry, consume far more water than data centers. The water used to maintain all U.S. golf courses for two days exceeds the total annual water consumption of our data centers worldwide; the total water consumption of golf courses is 200 times that of all our data centers.

Second, air pollution. Outsiders often cite theoretical emissions data from backup diesel generators, but in actual operations, because generators are only started during rare outages or routine maintenance, their effective annual operating time is only about 10 hours, a utilization rate of just 0.1%, so they are not a regular pollution source at all.

Finally, residential electricity costs. America's grid infrastructure has long suffered from insufficient investment; even without AI, grid expansion would have been a problem sooner or later—AI has simply made this structural contradiction erupt earlier. In places such as Indiana, Georgia, and Mississippi, with the injection of our large-scale infrastructure investment, local overall electricity prices have actually stabilized or even declined after expansion. In addition, in St. Joseph County, Indiana alone, we paid about $3 billion in taxes, effectively giving back to local infrastructure construction. Development must take community interests into account, but we cannot stop eating for fear of choking.

The "dumbbell shape" of computing supply and demand: why not resell resources to the newcomers?

Q: Amazon's capital expenditure this year is as high as $220 billion, which works out to about $600 million invested per day. How do you judge whether such an investment scale is too large? Also, you previously mentioned that AI demand currently presents a "dumbbell shape." What exactly does that mean?

Jassy: In the early days of AWS, the business reached $60 million in annualized revenue in its first three years, and at the time we already felt that was an extremely fast rate of expansion. Yet in the first three years of this generative AI cycle, AWS's AI-related business alone surpassed $25 billion in annualized revenue—a completely different growth slope.

This surging demand presents a classic "dumbbell shape" in structure:

At one end of the dumbbell are a very small number of top-tier model labs and a handful of super apps, whose consumption of compute and power is unprecedented; at the other end are lightweight scenarios that a broad range of enterprises are now deploying, mainly for point-solution cost reduction such as customer service automation, anti-fraud, or business process automation.

And the "middle section of the dumbbell"—the largest and most value-concentrated part of the entire business world, namely the massive production systems that major enterprises run in the cloud—has for the most part not yet substantively connected to large-model inference. Over the next 5 to 7 years, even if not the majority, a considerable number of enterprise applications will deeply integrate inference capabilities, and the explosion in this core range has only just begun.

As for whether this capital expenditure is excessive, the facts are the most powerful proof: even if we invest $220 billion this year, we still cannot meet all customer demand. Computing capacity for 2026 is already sold out, 2027 is basically fully booked, and a large portion of 2028 quotas has already been locked in.

In terms of the structure of capital expenditure, we clearly distinguish two types of assets:

One is the physical data center facility (land, buildings, and power connections), which requires about two years of lead time to build but has an asset useful life of up to 30 years;

The other is the servers and networking equipment inside the facility, which require only a few months of procurement lead time, typically pay back in about 3 years, have a hardware life of 5 to 6 years, and customers generally sign 5-year contracts.

This asset cycle means that although large-scale construction in the early stage will pressure short-term free cash flow, our risk control on server investment is very precise—once demand shifts, we can quickly scale back equipment purchases, while the facility infrastructure, as a long-lived durable asset, retains extremely high value.

Q: High-bandwidth memory, chips, and racks in the supply chain are now extremely scarce, and top AI labs are willing to pay sky-high prices for power and computing resources. Faced with this premium temptation, do you feel a business impulse: to take back traditional resources long rented to old customers and resell them at high prices to these AI companies that spare no cost?

Jassy: From a pure financial arbitrage perspective, that logic seems reasonable, but operationally we would never do it.

Indeed, whether high-bandwidth memory or general-purpose server memory, the supply chain is extremely tight right now, and resource allocation is testing every cloud provider. But in AWS's values, business reputation always comes before short-term gain.

During the pandemic, many peers thought about how to pull in as much revenue as possible each quarter, while we did the opposite: we proactively sent technical architects to help customers optimize their cloud architecture and guide them on how to scale down and reduce costs. That lowered our revenue growth in the short term, but it kept customers alive during a difficult period and built long-term bonds of trust.

With traditional core customers, the logic is the same. These customers' business lifelines have run stably on AWS for more than a decade, even twenty years. If we artificially squeezed old customers' resources or forcibly raised prices because of the immediate compute premium, we would be undermining the very foundation of the business.

Our job is to break supply bottlenecks as much as possible—safeguarding steady expansion for old customers while opening up incremental space for new AI compute. This tight supply chain situation will likely persist for several years; even if we invest much more capital expenditure, it would not be poor capital allocation, because market demand is far from fully satisfied.

Custom chips and inference costs: the key to determining how fast enterprises can deploy

Q: To address supply chain bottlenecks, major cloud providers are all pushing custom hardware. Amazon acquired the chip design company Annapurna Labs back in 2015, when the industry had almost no awareness of large models. What was the original purpose of that acquisition more than a decade ago? And how did it evolve into a core moat for AWS today?

Jassy: Acquiring Annapurna is one of the most successful acquisitions in Amazon's history.

But frankly, doing that deal in 2015 was absolutely not for today's AI large models. The core pain point at the time was this: as cloud computing scaled up, a considerable portion of server CPU compute was being used for management tasks such as virtualization, network transmission, storage management, and low-level security. We wanted to free up all of the server's compute for customers' business workloads rather than consuming it on system operations.

At the time, Annapurna was jointly developing with us a dedicated chip-based network card to offload these low-level overheads from the main CPU. After the acquisition, this technology evolved into today's Nitro system, deployed in every AWS server.

Through the collaboration, we validated the team's outstanding chip-making capability. At the time, the server CPU market was almost monopolized by Intel, and in the absence of competitors, the dominant supplier typically lacked the incentive to push price-performance to the limit. To help customers further reduce cloud migration costs, we had the Annapurna team develop Graviton, a server CPU chip based on the ARM architecture. Today, 99% of AWS's top 1,000 core customers are heavy users of Graviton, which offers about 40% better price-performance than same-generation x86 processors.

Back to the "dumbbell middle" mentioned earlier: why are the core production systems of massive numbers of traditional enterprises so slow to adopt AI? Talent and organizational capability are one factor, but a large part of the obstacle is that the economics don't work out—inference costs are too high. If the marginal cost of an application calling a model doesn't come down, the ROI for enterprises rewriting legacy systems will be very low. Inference costs must now drop by another order of magnitude before a true explosion of enterprise-grade applications arrives.

This is the core reason we developed our own AI chip, Trainium. Compared with similar hardware, the latest generation of Trainium likewise maintains a 30% to 40% advantage in price-performance. Only by driving down inference costs can AI applications go from lab toys to ubiquitous production tools.

At the same time, the underlying logic of allying with both Anthropic and OpenAI

Q: Beyond underlying hardware, Amazon has also been making frequent moves in ecosystem partnerships. You invested tens of billions of dollars in Anthropic, and recently deepened a large-scale partnership with OpenAI. But to the capital markets, the model of cloud vendors acting as both investors and compute suppliers is controversial. What is the business logic behind Amazon's two heavyweight partnerships?

Jassy: These two companies have among the world's top engineering teams today, and in terms of business growth speed they have broken records in the software industry.

Our investments and partnerships are based on three clear dimensions: first, we are confident that AI represents a profound industrial transformation; second, the large-model ecosystem will not evolve into a monopoly by a single player—both of these teams will be winners in the long run; third, investment must rest on real, deep technical collaboration.

With Anthropic, the partnership has lasted longer. Massive numbers of enterprise customers choose to call the Claude series models through our managed platform Bedrock. More importantly, Anthropic is also a core R&D partner for our self-developed Trainium chips; a large number of their frontier training and inference workloads have long run on Trainium clusters, and this software-hardware collaboration benefits both sides greatly.

On the OpenAI side, the trigger for the partnership came entirely from strong customer demand. In the past, limited by commercial terms, it was very difficult for enterprises to directly call OpenAI's frontier models within AWS's native secure environment. After multiple rounds of discussions, we not only successfully brought OpenAI's models into Bedrock, but also jointly developed a "stateful runtime environment."

Future AI agents must have long-term memory and state-retention capabilities, and this technology forms the foundational layer for next-generation agent applications. At the same time, OpenAI has also made a long-term commitment to purchase several gigawatts of Trainium compute resources. This kind of partnership—which both meets the needs of core enterprise customers and finds a deployment scenario for our self-developed hardware—is strategically very clear.

Six people rewrote a system in 76 days: AI reshapes team organization models

Q: Let's look at how AI is transforming the organizational form of large enterprises. You mentioned that the Bedrock team used only 6 engineers and took 76 days to overturn and rewrite the core inference engine, whereas the traditional model would require 40 people for a full year. Amazon famously introduced the "two-pizza team" concept back in the day. Today, with AI-assisted programming and the spread of agents, how do you view the evolution of team structures in large companies? Has the fine-grained division of labor of the past become obsolete?

Jassy: The Bedrock team's rebuild caused a big stir inside Amazon. We have repeatedly told this case across the whole company, precisely to break the organizational inertia that large enterprises develop as they scale.

I've been at Amazon for almost 30 years. When I first joined, I was an assistant product manager, and I alone was responsible for 5 products. Yet as the company grew larger, the organization inevitably moved toward over-specialization: a product manager often was responsible for only an extremely narrow module under a very large system, for example solely for pricing strategy or a specific interaction. Everyone gradually developed a kind of habitual thinking—that for a large company to advance a business, it must rely on lengthy processes and huge teams.

But in the AI era, this logic is being rapidly broken. Last week, for a brand-new agent concept for the e-commerce business, we finalized the plan on Thursday, and with only four or five engineers using AI development tools, we delivered by Sunday a prototype that was 80% complete and could actually run. That would have been unimaginable in the past.

I think this evolution will bring several certain changes to corporate organizations:

First, the number of people inside large enterprises doing routine, repetitive work will definitely continue to shrink in the future. Whether it's basic code writing, day-to-day customer service, basic data analysis, or preliminary legal review, the old model of relying on sheer numbers is unsustainable. Although emerging businesses will create new roles, the reduction of traditional jobs is irreversible.

Second, teams will become more compact, and the value of versatile talent will be amplified. AI can fill in 80% of the foundational capabilities across technical and business functions, which means members are no longer constrained by rigid functional divisions and can more independently drive end-to-end business implementation.

Finally, enterprise software architectures must be fundamentally reworked for agents. Today's AI-native startups design systems for automated workflows from day one; but if mature enterprises still cling to decades-old legacy codebases and manual audit processes, they simply cannot keep up with the pace of iteration. Enterprises must make their systems capable of being quickly intervened in, modified, and deployed by agents.

Q: What new requirements does this rapid evolution toward automation place on system security? The industry has recently been keen to discuss models' "existential risk"—does that distract public attention from real technological hazards?

Jassy: It does. Lately there have been frequent, somewhat science-fiction-flavored extreme concerns in public discourse, such as whether malicious AI systems might spiral out of control and take over infrastructure. These theoretical risks do need guarding against, but frankly, government, industry, and academia have enough time to build rules and protective nets.

The real hazard that is truly lethal today but severely under-discussed is actually cybersecurity.

Large models can find far more system vulnerabilities and code defects than before—many that human auditors never discovered. In the traditional model, going from discovering a vulnerability, assessing the threat, developing a patch, to coordinated deployment across systems often drags on for a long time; deployment alone is often blocked by all kinds of dependencies. But in AI-enabled cyber offense and defense, responding to imminent threats must be measured in minutes.

In such an environment, enterprises must rework their code architecture so systems have the ability to "have agents automatically apply patches and complete hot deployment within minutes." If the defense cadence remains stuck at the manual scheduling stage of the past, systems will have almost no ability to fight back against future attacks. We have also launched an automated security service for customers called Continuum, dedicated to finding vulnerabilities and generating patches. Security work is tedious and dull, but it is a life-or-death line of defense.

Rebuilding Alexa: Mature products can't just get incremental patches

Q: The industry currently has roughly a few approaches to using AI: one is cutting costs and headcount through AI; another is developing unprecedented new species; and the third is what you're trying on Alexa—tearing down and rewriting an entire mature product with hundreds of millions of users. How much resistance within the organization do you face when overturning your own past achievements?

A generational leap in underlying technology means patching has no way out. Alexa's technical approach was very advanced when it launched, but the emergence of large language models has made that old core completely obsolete. We had to hollow out and replace Alexa's underlying architecture entirely, even if the process is extremely grueling.

The biggest resistance to this kind of rebuild often comes from inside the organization. The original team poured years of effort into the old architecture, overcame countless technical hurdles, and has deep emotional attachment to the existing approach. When you ask everyone to tear it down and start over, the most common pushback is: "This approach proved effective in the past, and we stepped in those pits years ago with the new ideas you're proposing."

Breaking a team's established mindset takes a great deal of time and managerial energy.

And harder than rebuilding Alexa is rebuilding those core businesses that are currently running well and highly profitable—such as our retail platform, the Prime Video streaming service, and the digital music business. On the surface these systems show no crisis at all, but we still pull independent teams internally to try incubating entirely new versions of these businesses from scratch, based purely on AI-native architecture.

We can't get every approach right on day one, but this kind of self-reinvention is a pain we must endure. If we don't proactively disrupt ourselves with the new paradigm, we'll be disrupted by outside emerging teams. In the history of technology, whether it was the migration from on-premises IDCs to cloud computing back then or the AI-oriented architecture rewrites today, it is essentially all highly challenging organizational change management.

Advice for young people: The lower the barrier to tools, the more valuable critical thinking becomes

Q: As AI tools dramatically lower the technical barrier, many young people are starting to feel lost. If you could go back to being 18, standing here in 2026, what advice would you give yourself?

Jassy: I have two children, and they entered college or the workforce right as this AI wave erupted. I realized then: any idea of trying to keep AI tools out of education and work is unrealistic—the wheel of the times has already turned.

Amazon has always maintained an extremely strict "written memo" culture; we don't use slide presentations, we rely entirely on writing detailed six-page documents. Over the past year, many employees have started using AI to assist with writing. Frankly, you can tell after reading two paragraphs whether a document was entirely ghostwritten by AI—the wording is standard, but the ideas lack the logically interlocking soul.

If I were advising an 18-year-old, I would stress two things:

First, don't avoid it—embrace it fully.

You must treat AI tools as a standard means of daily work and learning. Go fully try things beyond the boundaries of your own abilities. My two children have no professional computer engineering training, but they can now skillfully use AI tools to independently develop complete applications. Treat it as an amplifier of intelligence; don't develop a resistance to it.

Second, never give up critical thinking and deep judgment.

It's like analyzing cases at business school: in the past, professors would assign complex business cases and lengthy data tables, and there were always some students unwilling to do solid reading and logical reasoning, who couldn't offer substantive insights in class discussion.

But after that, how do you connect scattered business information into effective insight? How do you judge non-consensus business opportunities? How do you design products that move people? How do you make decisions amid chaos? This work still relies on human intuition, empathy, the courage to trial and error, and deep critical thinking.

The lower the barrier to tools, the higher the value of independent thinking.

With a business span from satellites to retail, how does Amazon avoid overextending itself?

Q: From the earliest online bookstore to today's e-commerce, hyperscale cloud infrastructure, underlying chips, warehouse robots, digital health, and the Amazon Leo space broadband project deploying thousands of low-Earth-orbit satellites, Amazon's battle lines are extremely long. Many giants in business history have fallen into trouble because of over-diversification. How do you ensure that Amazon, while involved in so many frontier areas at the same time, is not overstretched?

Jassy: We believe that human society is currently at the intersection of three major technological transformations: artificial intelligence, robotics and automation, and low-Earth-orbit space infrastructure.

Each of these three transformations has the potential to reshape the operating rules of many industries in the coming decades. And Amazon is not jumping on these trends; our original core businesses are rooted at the center of these tracks:

In retail, we use robots extensively; in AI, we both make it easier for internal and external developers to build successful AI applications and agents, and are reshaping all consumer-facing experiences; and in space infrastructure, our low-Earth-orbit satellite project aims to provide broadband connectivity to 500 million households worldwide that lack stable network coverage.

To avoid losing control because the business front is too long, two things are key: organizational design principles and long-term strategic resolve.

Amazon never lets the same management team oversee businesses with vastly different spans. We have long adhered to a "single-threaded leadership" accountability system—for every major new strategic business, we select capable leaders with pioneering ability and build an independent, closed-loop team. The heads of these teams wake up every day with the sole mission of going deep in that one vertical and not being distracted by the noise of other day-to-day businesses.

Some companies tend to stick to their existing core base and maximize the profits of a single business. That is a very successful strategic path; but Amazon's DNA is different. As long as a market is large enough, there are unmet pain points, we have the ability to significantly improve the experience through technology, and we can ultimately generate healthy long-term cash flow, we will resolutely make strategic investments.

Focusing on doing the existing core business well and maintaining the ambition to pioneer unknown areas are not contradictory.

Source: The a16z Show, Ben Horowitz and Erik Torenberg in conversation with Andy Jassy, aired October 23, 2026, https://www.youtube.com/watch?v=-YMMRQ1qscIAI Industry Signal Channel: Frontier Technology This content comes from the internet and the original link. The views represent only the author and do not represent the position of Huxiu. If there is a copyright issue, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is on the Huxiu APP [ ]