Five Things to Watch in AI for the Rest of 2026

The questions that get answered between now and December

Written 3 August 2026. Every item below has a dated trigger in the next five months rather than a vague trajectory. Figures reflect reporting available at the time of writing and several are contested; the article says so where that is the case.

---

Forecasting AI is mostly a genre of confident nonsense, because the honest answer to most questions about it is "nobody knows." So this piece takes a narrower approach. Rather than predicting where the technology goes, it identifies five things that are scheduled to resolve, or come close to resolving, before the end of the year. Each has a date attached, a specific question underneath it, and an observable signal that would tell you which way it went.

The first half of 2026 delivered a $122 billion private funding round, a $200 billion quarter at Amazon, a 19-day government-ordered shutdown of two frontier models, and the first amendment to the EU AI Act. The second half looks busier.

---

1. Public markets finally get to price a frontier AI lab

The date: October, possibly slipping.

Anthropic filed a confidential S-1 with the SEC on 1 June 2026, with Goldman Sachs, Morgan Stanley and JPMorgan leading. Confidential review for a company this size typically runs three to six months, which points to a public S-1 around September, a roadshow through late September and early October, and a first trade in October. [1][2] (Disclosure: Anthropic makes Claude, the assistant that researched and drafted this article.)

OpenAI filed a week later, on 8 June, then immediately took the pressure off, saying publicly that it had not decided on timing. Bloomberg reported that OpenAI's leadership expects Anthropic to list first, and CFO Sarah Friar has reportedly argued internally for 2027, citing cash burn and the scale of compute commitments. [2]

Why this matters more than a typical listing: there is currently no public market price for frontier AI model revenue. Everything, private rounds, secondary trades, the valuations of every AI startup downstream, is being marked against other private rounds. The first frontier lab to list establishes a public multiple that every banker's model uses as a denominator from that day forward. If a lab prices at 20x forward revenue, that becomes the anchor. If it prices at 8x, a great deal of private paper gets remarked.

Cerebras is the early read. The AI chipmaker listed on Nasdaq on 14 May 2026, opened far above its offer price and closed up roughly 68% on debut, then gave back about 10% the following day. [3][4] Appetite exists. Whether it survives contact with an S-1 that discloses a loss line is a different question.

What to watch: the public S-1 release, whenever it lands. The risk factors and the audited revenue figures will settle several arguments that run on estimates today, including whether the widely cited run-rate numbers hold up against trailing GAAP revenue. Run rate is last-month-annualised; it is not the same object.

What would change the picture: a delay past October in either filing, or a first-day close meaningfully below the last private round. Either would suggest institutional investors are pricing the loss line rather than the growth line.

---

2. Power, not chips, decides who actually ships

The date: continuous, but visible in Q3 and Q4 capacity announcements.

The bottleneck has moved. For most of 2023 and 2024 the constraint was packaging capacity and GPU allocation. It is now the substation. Grid interconnection in major US and European markets runs 24 to 36 months, high-voltage transformers carry lead times measured in years, and US interconnection queues have swollen past 2,100 gigawatts, which exceeds total grid capacity. [5][6]

The consequence is already quantified. Industry analysis projects that 30% to 50% of planned 2026 data centre capacity will slip to 2028 because of power and construction constraints, and Sightline Climate's February 2026 work found nearly half of global data centre projects due for completion this year facing power-related delays. [5][7] Morgan Stanley's modelling points to a 49 gigawatt shortfall in the US alone through 2028. [7]

Silicon is constrained too, in a specific place. HBM capacity for 2026 was preallocated in its entirety by SK Hynix, Micron and Samsung, and HBM gross margins have reportedly run at 60% to 70%, well above standard DRAM. [6] Memory pricing has already shown up in capex guidance: Microsoft's CFO attributed a chunk of the company's raised 2026 capital spending to rising memory and component costs. [8]

This is the least glamorous item on the list and probably the most determinative. Capital is not the constraint. Physics and permitting are.

What to watch: the gap between announced gigawatts and energised gigawatts. Companies announce projects in gigawatts and deliver them in megawatts, on timelines that keep extending. Watch for capex guidance that holds steady while delivery dates move right, which is the polite way an infrastructure boom decelerates.

What would change the picture: meaningful movement on behind-the-meter generation. If on-site gas or nuclear arrangements start energising at scale rather than being announced, the constraint loosens faster than the interconnection queue suggests.

---

3. The Rubin and MI450 generation lands, and the depreciation clock restarts

The date: Q3 shipments, Q4 volume ramp.

Nvidia confirmed that Vera Rubin production shipments begin in Q3 2026 with volume ramping in Q4, and said in July that the Vera Rubin NVL72 platform had entered production ramp with early rack-scale systems running at CoreWeave, Google Cloud, Azure and Oracle Cloud. [9][10] Rubin moves to TSMC 3nm, a chiplet design, and HBM4.

AMD arrives in the same window. The company began sampling MI450 to lead customers and is on track to ramp Helios rack production in the second half of the year, with Lisa Su noting on the Q1 2026 call that lead customer forecasts had come in above initial plans. [11] Oracle committed to 50,000 MI450-series GPUs starting in Q3 2026. The first gigawatt of OpenAI's 6 gigawatt AMD commitment also begins deploying in the second half. [12]

On paper AMD has a memory advantage, 432GB of HBM4 against Rubin's 288GB, while Nvidia counters on interconnect bandwidth. [13][14] The spec comparison is less interesting than the deployment question, which is whether AMD converts its two very large publicly disclosed commitments into installed capacity on schedule.

The second-order effect is the one to think about. Every new generation that delivers a large step in performance per dollar shortens the economic life of the previous generation, which is precisely the assumption underpinning the depreciation schedules that hyperscaler earnings depend on. Rubin shipping on time strengthens the argument that GPU useful life is closer to three years than six. That argument, if it wins, has consequences for reported profit across the sector.

What to watch: whether any hyperscaler shortens its stated server useful life in Q3 or Q4 disclosures. Amazon already cut from six years to five in early 2025 citing the pace of AI development; Meta went the other way. [15] A second major shortening would be the most concrete validation the depreciation critique could get.

What would change the picture: Rubin slipping. There were reports in 2025 of a possible delay tied to a second tape-out, which Nvidia denied. [16] A slip would relieve depreciation pressure and hand AMD a wider opening simultaneously.

---

4. Agents have to show up in somebody else's earnings

The date: Q3 earnings in October, Q4 guidance in January.

The enterprise agent data is currently a mess, and that is the finding rather than a complaint about sourcing. Depending on which 2026 survey you read, the share of enterprises running at least one agent in production is 52%, 31%, or 14%. [17][18][19] Reported median ROI on production deployments runs as high as 171% globally. [18] Meanwhile Gartner projects that 40% of enterprise applications will embed task-specific agents by the end of 2026, up from under 5% a year earlier, and separately warns that over 40% of agentic AI projects are at risk of cancellation by 2027. [19][20]

Both of those Gartner numbers cannot describe the same world comfortably. The reconciliation is probably that adoption is broad and shallow: many agents deployed, few of them load-bearing.

The reason this matters for the rest of 2026 is that the single most damaging data point in the AI bear case has been the repeated finding that a large majority of enterprise generative AI deployments show no measurable return. [21] The corresponding bull signal is specific and testable, and it is not another vendor survey. It is a non-technology company in the S&P 500 attributing concrete margin expansion to AI on an earnings call, with a number attached, that survives analyst questioning.

That has not happened at scale. October is the next opportunity.

What to watch: Q3 earnings calls outside the technology sector. Banks and insurers are furthest along on agent deployment and have the cleanest measurement of the relevant workflows, so financial services is the sector most likely to produce a credible first data point.

What would change the picture: a wave of quiet cancellations. Gartner's cancellation warning is a forecast about 2027, but the decisions that produce it get made during budget season, which is now.

---

5. Governments become an operating dependency, not a compliance line item

The dates: 2 December 2026, and whenever the next directive arrives.

Two things happened in the first half of the year that changed the category of political risk in AI from "future regulation" to "operational dependency."

The first was the EU AI Act's amendment. On 29 June 2026 the Council of the EU gave final approval to the Digital Omnibus on AI, after the Parliament endorsed it on 16 June by 423 votes to 57. High-risk obligations for standalone Annex III systems moved from 2 August 2026 to 2 December 2027, and for AI embedded in Annex I regulated products to 2 August 2028. [22][23][24]

The relief is real but narrower than the headlines suggested. The Article 50 transparency obligations were not deferred and took effect on 2 August 2026: chatbot disclosure, and marking of synthetic audio, image, video and text. And a live date remains inside this year. On 2 December 2026, Article 50(2) transparency requirements extend to systems already on the market before August, and the new prohibited practices take effect, including a new prohibition on AI-generated non-consensual intimate imagery and CSAM added to Article 5. [24][25]

The second event was sharper. On 12 June 2026, three days after Anthropic released Claude Fable 5 and Mythos 5, the Commerce Department's Bureau of Industry and Security issued an export control directive requiring the company to suspend all access to both models by any foreign national, anywhere, including its own non-citizen employees. Unable to screen users by nationality, Anthropic disabled the models globally. They stayed down for 19 days. Commerce lifted the controls on 30 June and access was restored on 1 July, after the company agreed to proactively detect and address security risks, work with the government on standards for future models, and report malicious activity. [26][27][28][29]

Both sides disputed the substance at the time, and readers should treat the technical claims as contested rather than settled. What is not in dispute is the precedent: the US government demonstrated that it can and will use export control authority to switch off a commercially deployed frontier model, on short notice, worldwide. [30]

Set that alongside the chip picture, where Commerce shifted to case-by-case licensing for H200 sales to China in January 2026 with conditions attached, then issued weekend guidance in late May clarifying that license requirements apply to entities headquartered in China even when the entities sit outside it. [31][32] The direction of travel is toward more discretionary, faster-moving, less predictable intervention.

What to watch: whether the Fable and Mythos episode turns out to be a one-off or the first instance of a repeatable procedure. The relevant signal is whether any formal standard emerges for pre-deployment government review of frontier models, or whether it stays ad hoc.

What would change the picture: a published framework. Ad hoc directives are the riskiest state of affairs for everyone, including the government. A codified process, even a demanding one, would be a meaningful change in the operating environment.

---

What this list leaves out, on purpose

Three things that will get a lot of coverage and probably will not resolve this year.

AGI timelines. No date, no falsifiable trigger, no agreed definition. Skip.

The bubble question. Whether AI capital markets are in a bubble will not be settled in five months. What will be visible is the divergence already underway between credit markets, which have been widening spreads on the large AI borrowers, and equity markets, which have not repriced to match. That divergence is worth tracking, but it is a slow signal, not a Q4 event.

Model releases. New frontier models will ship. They almost always beat their predecessors on benchmarks, and the benchmark result almost never tells you what changed commercially. The commercially interesting question is not which model is best but whether inference costs fall fast enough to make the unit economics work, and that shows up in gross margin disclosures rather than in launch posts.

---

The through-line

Four of these five items are about physical and institutional constraints rather than about capability. Grid interconnection queues, HBM allocation, audited financial statements, government directives. That is the shape of a technology moving out of its research phase and into its infrastructure phase, where the binding constraints stop being intellectual and start being logistical and legal.

The interesting version of the question for the rest of 2026 is not how capable the models get. It is whether the surrounding system, the power grid, the debt markets, the public equity markets, the regulators, can absorb the pace at which they are being deployed. Those systems all have their own clock speeds, and none of them run at the speed of a model release.

---

References

1. Anthropic confidential S-1 filing and underwriter reporting, June 2026, summarised in Axis Intelligence, "Anthropic Statistics 2026." https://axis-intelligence.com/anthropic-statistics/
2. INDmoney, "Anthropic vs OpenAI IPO Date: Who Will List First?" June 29, 2026, on filing dates, underwriters and reported internal timing disagreements. https://www.indmoney.com/blog/us-stocks/anthropic-openai-ipo-date-valuation-risks
3. Reuters reporting on the Cerebras Systems debut, May 14, 2026, summarised in BeInCrypto, "Top AI IPOs To Watch in 2026." https://beincrypto.com/learn/top-ai-ipos/
4. TradingKey, "Cerebras Surges 68% on Debut to Kick Off AI Listing Wave," May 2026. https://www.tradingkey.com/analysis/stocks/us-stocks/261898821-cbrs-cerebras-spacex-openai-anthropic-ipo-tradingkey
5. Manufacturing Dive / Omdia, "The great data center delay: Why your AI chips are stuck in 2026," April 20, 2026. https://www.manufacturingdive.com/news/opinion-omdia-ai-semiconductor-chip-scarcity/817172/
6. Accuris, "How AI Data Centers Are Reshaping Electronic Component Supply in 2026," May 27, 2026, on capacity slippage and HBM preallocation. https://accuristech.com/blog/ai-data-center-electronic-component-supply/
7. Tech Insider, "The AI Data Center Power Crisis," June 4, 2026, citing Uptime Institute, Sightline Climate and Morgan Stanley. https://tech-insider.org/ai-data-center-power-crisis-2026/
8. Tom's Hardware, summarising Financial Times compilation of Q1 2026 hyperscaler capex guidance and Microsoft CFO commentary on memory costs, April 30, 2026. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
9. Wccftech, "NVIDIA Confirms Vera Rubin Launch In Q3 With Volume Ramp by Q4," May 21, 2026. https://wccftech.com/nvidia-confirms-vera-rubin-launch-in-q3-volume-ramp-q4-blackwell-continues-to-see-massive-demand/
10. Moomoo Community, on Nvidia's July 21, 2026 confirmation of Vera Rubin NVL72 production ramp and named partner systems. https://www.moomoo.com/community/feed/nvidia-s-vera-rubin-is-in-full-production-mapping-the-116963864019350
11. Wccftech, "AMD Has Begun Sampling MI450 GPUs," May 5, 2026, quoting Lisa Su on the Q1 2026 earnings call. https://wccftech.com/amd-sampling-mi450-gpus-engaged-with-customers-on-mi500-largest-ai-deployments-are-for-inference/
12. Silicon Report, "AMD MI455X vs Nvidia Vera Rubin," July 3, 2026, on the Oracle commitment and OpenAI deployment timing. https://www.siliconreport.com/amd-mi455x-vs-nvidia-vera-rubin-bd525bb4
13. Introl, "NVIDIA Rubin Enters Full Production," January 2026, on the memory and interconnect comparison. https://introl.com/blog/nvidia-rubin-full-production-ces-2026-ai-infrastructure
14. SemiAnalysis, "Vera Rubin: Extreme Co-Design," February 2026, on HBM4 bandwidth targets and supplier constraints. https://newsletter.semianalysis.com/p/vera-rubin-extreme-co-design-an-evolution
15. Dave Friedman, "Neoclouds Hold More Than $20 Billion in GPU-Backed Debt," February 2026, documenting hyperscaler useful-life changes. https://davefriedman.substack.com/p/neoclouds-hold-more-than-20-billion
16. TrendForce, "NVIDIA Reportedly Denies Rubin Delay as AMD MI450 Spurs Redesign Speculation," August 2025. https://www.trendforce.com/news/2025/08/14/news-nvidia-reportedly-denies-rubin-delay-as-amd-mi450-spurs-redesign-speculation/
17. Anandbg, summarising Google Cloud's ROI study of 3,466 senior leaders across 24 countries. https://anandbg.com/blog/enterprise-ai-agents-the-2026-strategic-playbook
18. eCorp IT, "7 enterprise AI agent use cases in production (2026)," June 26, 2026, on compiled ROI survey data and McKinsey scaling figures. https://ecorpit.com/enterprise-ai-agents-production-use-cases-2026/
19. Paul Okhrem, "Enterprise AI Agents Adoption Statistics 2026," citing S&P Global Market Intelligence, McKinsey, BCG, Forrester and Gartner. https://paul-okhrem.com/enterprise-ai-agents-statistics-2026/
20. Cygnet, "AI Agents for Enterprise Deployment: Best Practices 2026," on the Gartner 40% embedding projection and ungoverned sprawl. https://www.cygnet.one/feeds/blog/ai-agents-enterprise-deployment-november-2025
21. IntuitionLabs, "AI Bubble vs. Dot-com Bubble: A Data-Driven Comparison," including the MIT enterprise ROI finding. https://intuitionlabs.ai/articles/ai-bubble-vs-dot-com-comparison
22. Freshfields, "EU AI Act unpacked #34: The final Digital Omnibus on AI," June 2026. https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-34-the-final-digital-omnibus-on-ai-key-amendments-to-the-a-102nber
23. Secure Privacy, "EU AI Act Digital Omnibus: The New High-Risk AI Deadlines After Council Approval," July 2026, on the Parliament vote count and Council approval date. https://secureprivacy.ai/blog/eu-ai-act-digital-omnibus-the-new-high-risk-ai-deadlines-after-council-approval
24. Winston Taylor, "AI Act rules on high-risk AI delayed as AI Digital Omnibus agreed," setting out the revised compliance timetable including 2 December 2026. https://www.winstontaylor.com/insights/ai-act-rules-on-high-risk-ai-delayed-as-ai-digital-omnibus-agreed
25. Gibson Dunn, "EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines and Other Key Changes," May 2026, on the new Article 5 prohibitions. https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
26. Anthropic, "Statement on the US government directive to suspend access to Fable 5 and Mythos 5," June 12, 2026. https://www.anthropic.com/news/fable-mythos-access
27. Fortune, "Anthropic disables Fable and Mythos AI models following U.S. government export ban," June 13, 2026. https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/
28. CNBC, "Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5," June 30, 2026. https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html
29. Al Jazeera, "US lifts restrictions on Anthropic's powerful AI models Fable and Mythos," July 1, 2026, on the conditions in the Commerce letter. https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says
30. Tech Policy Press, "Did the US Government Just Set An AI Export Precedent by Blocking Mythos?" June 16, 2026. https://www.techpolicy.press/did-the-us-government-just-set-an-ai-export-precedent-by-blocking-mythos/
31. Tech Policy Press, "Technology Restrictions Have Become a Central Instrument of Economic Statecraft," April 2026, on the January 2026 BIS licensing shift and its conditions. https://www.techpolicy.press/technology-restrictions-have-become-a-central-instrument-of-economic-statecraft/
32. CNBC, "U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China," May 31, 2026. https://www.cnbc.com/2026/05/31/us-takes-step-to-halt-nvidia-ai-chip-shipments-to-chinese-firms-outside-china.html